Wireless LANs

INTERNET

IEEE International Conference on

Wireless LANs and Home Networks

Wireless LANs and Home Networks Connecting Offices and Homes

Wireless LANs and Home Networks Connecting Offices and Homes Proceedings of the International Conference on Wireless LANs and Home Networks Singapore

5-7 December 2001

Editor

Benny Bing Georgia Institute of Technology, USA

Technical Co-Sponsors

IEEE Communications Society

V f e World Scientific WW

New Jersey • London • Singapore • Hong Kong London'Singapore*

Published by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Farrer Road, Singapore 912805 USA office: Suite IB, 1060 Main Street, River Edge, NJ 07661 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library.

Cover design: Illustrates the importance of wireless LANs in the mobile Internet infrastructure.

IEEE International Conference on WIRELESS LANS AND HOME NETWORKS Copyright © 2001 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.

ISBN 981-02-4826-1

Printed in Singapore by World Scientific Printers (S) Pte Ltd

- Platinum Sponsor

Cisco S Y S T E M S

EMPOWERING THE INTERNET GENERATION"

Gold Sponsors^Sf ^

-VEIL. AT

Singapore Information Technology Federation

- Silver Sponsors

tntermec Supporting

3Com

Sniffer •• T E C H N O L O G I E S A Network

Associates

Business

Sponsors.

ABCHITECtS Of AN INTERNET WORID

N OICIA CONNECTING PEOPLE:';

ewcT

reefedge

onnOCO

~"~"

1 \. A T H EROS"

'embedded wireless devices inc

broadaand wireless networking on a c/i/p™

designed by adere sy

PREFACE

Wireless local area network technologies have changed dramatically in the past 2 years. In September 1999, the IEEE 802.11 Working Group for Wireless LANs finalized a 5 GHz standard (approved as IEEE 802.11a) that will support wireless data rates ranging from 6 to 54 Mbit/s. In an extension to the 2.4 GHz IEEE 802.11 standard (approved as IEEE 802.11b on September 1999), information can be transferred at data rates of up to 11 Mbit/s. These new developments have significant implications in that the standards are now more than capable of supporting streaming audio and video traffic needed for mobile multimedia applications. However, since these wireless LAN standards (802.11a and 802.11b) operate on different frequency bands (2.4 and 5 GHz) with differing rates, it is unclear which standard will eventually prevail. With new wireless personal devices involving Bluetooth starting to gain traction commercially, the wireless world continues to create difficult options for both end users and service providers. The 2001 International Conference on Wireless LANs and Home Networks showcases some of the world's most dynamic presenters, including Dr. Leonard Keinrock (Inventor of Internet Technology) as well as leading experts from 20 countries who presented the latest technology breakthroughs. This book is a collection of technical papers that were presented at the conference. It comprises 32 high-quality papers that were carefully selected from more than 100 submissions, many of which were worthy of publication. Among these accepted papers, authors of 5 best papers were invited to contribute articles for the IEEE Wireless Communications Magazine (formerly IEEE Personal Communications Magazine), a reknown and highly respected technical journal. In addition, a collection of white papers and tutorials from different companies is published in a separate volume. I am grateful to the program committee, the sponsors and the presenters for their time and tireless efforts in making this conference a reality and a success. Sincere thanks also go to the World Scientific staff namely Lim Sook Cheng, Yolande Koh and Chelsea Chin for their diligence in completing this project in a timely manner. I hope you will enjoy reading many of the ground-breaking innovations in this volume and look forward to your support in the next conference. Benny Bing Singapore, 2001 www.icwlhn.org vii

TECHNICAL PROGRAM COMMITTEE

Lek Ariyavisitakul, Home Wireless Networks, USA. Ben Arzine, British Telecom, UK. Ender Ayanoglu, Cisco Systems, USA. Yeheskel Bar-Ness, New Jersey Institute of Technology, USA. John Barr, Motorola, USA. Steve Bell, Agilent Interoperability Certification Labs, USA. Justin Chuang, AT&T Research, USA. David Cohen, 3Com, USA. Greg Ennis, Symbol Technologies, USA. David Everitt, Swinburne University of Technology, Australia. Laurent Frelechoux, IBM Research, Switzerland. Alex German, Panasonic Research, USA. Nada Golmie, National Institute of Standards and Technology, USA Lon Gowen, Mitre Corporation, USA. Bob Heile, Consultant, USA. Chin-Lin I, AT&T Research, USA. Konosuke Kawashima, NTT Advanced Technology, Japan. Parviz Kermani, IBM Research, USA. Leonard Kleinrock, University of California at Los Angeles, USA. Victor Li, University of Hong Kong, China. Pascal Lorenz, University of Haute Alsace, France. Teresa Meng, Stanford University, USA. Jouni Mikkonen, Nokia, Finland. Hiroyuki Morikawa, University of Tokyo, Japan. Guy Pujolle, University of Paris, France. Mike Sheppard, Ericsson, USA. Matthew Shoemake, Texas Instruments, USA. Roj Snellman, Intersil, USA. Peter Steenkiste, Carnegie Mellon University, USA. Richard van Nee, Lucent Technologies, The Netherlands. Sergio Verdu, Princeton University, USA. Naoaki Yamanaka, NTT Network Service Systems, Japan. Rodger Ziemer, National Science Foundation, USA.

IX

LIST OF REVIEWERS

In addition to the program committee, the efforts of the following reviewers are gratefully acknowledged. A. Ahmad, DePaul University, USA. S. Buchegger, IBM Research, Switzerland. Y. Cao, University of Hong Kong, China. K. Chan, University of Hong Kong, China. W. Hirt, IBM Research, Switzerland. Z. Lei, University of Hong Kong, China. M. Katayama, NTT, Japan. K. Kawashima, NTT, Japan. M. Khan, Panasonic, USA. Z. Lu, University of Hong Kong, China. S. Okamoto, NTT, Japan. M. Osborne, IBM Research, Switzerland. D. Shimazaki, NTT, USA. K. Shiomoto, NTT, USA. Z. Zhang, University of Hong Kong, China. T. Znati, University of Pittsburgh, USA.

CONTENTS

Preface

vii

Part 1 Mobile Computing 802.11: No Strings Attached J. Fawcett (University of Cambridge, UK) and W. Sowerbutts (AT&T Laboratories Cambridge Ltd, UK)

3

Classroom in the Era of Ubiquitous Computing Smart Classroom C. Jiang, Y. Shi, G. Xu and W. Xie (Tsinghua University, China)

14

Hybrid Jini for Limited Devices V. Lenders, P. Huang and M. Muheim (ETH Zurich, Switzerland)

27

Part 2 Quality of Service and Wireless Internet A DiffServ-Based Classification Scheme for Internet Traffic Over Wireless Links D. Skyrianoglou, N. Passas and S. Kampouridou (University of Athens, Greece) A Management Entity for Improving Service Quality in Mobile Ad-Hoc Networks M Bechler, B. Hurler, V. Kahmann and L. Wolf (University of Karlsruhe, Germany) Integrating IPv4 and IPv6 in Wireless Networks E. Jamhour (Pontificia Universidade Catolica do Parana, Brazil)

XI

37

47

57

XII

Part 3 Error Control and Mobile Applications A Low Cost Error Control Scheme for Hard Real-time Messages on Wireless LANs J. Lee, Y. Lee and S. Kim (Ministry of Science and Technology, Korea) Bundles Replacement in Gateways B. Chen, K. Elhassioni (Rutgers University, USA) and I. Kamel (Panasonic Information and Technologies Laboratory, USA)

69

79

Part 4 Bluetooth and 802.11 Flash Notes Over Bluetooth Wireless Technology P. Jappinen and J. Porras (Lappeenranta University of Technology, Finland) Performance Enhancements to the IEEE 802.1 lb Standard M. Shoemake (Texas Instruments, USA)

91

100

Part 5 Network Security A User Authentication System for Secure Wireless Communication N. Yamai (Okayama University, Japan) H. Ishibashi, K. Abe, T. Matsuura (Osaka City University, Japan) H. Morishita and T. Mori (Stella Craft, Inc., Japan)

113

Design, Implementation and Evaluation of Bluetooth Security J.-Z. Sun, D. Howie, A. Koivisto and J. Sauvola (University of Oulu, Finland)

121

Your 802.11 Wireless Network has No Clothes W. Arbaugh, N. Shankar and J. Wan (University of Maryland, USA)

131

XIII

Part 6 Power Control and Performance Evaluation Vertical Optimization of Data Transmission for Energy Aware Mobile Devices K. Dombrowski, M. Methfessel, P. Langendorfer, H. Frankenfeldt, I. Babanskaja, I. Matthaei and R. Kraemer (IHP, Germany) RADIOSCAPE: System Design Tool for Indoor Wireless Communications via the Internet Y. Watanabe, H. Furukawa, K. Okanoue and S. Yamazaki (NEC Networking Research Labs, Japan) Capture Effect in IEEE 802.11 Wireless LANs Z Hadzi-Velkov andB. Spasenovski (Sts. Cyril and Methodius University, Macedonia)

145

154

164

Part 7 Medium Access Control Optimizing the Polling Sequence in Embedded Round Robin WLANs L. Andrew andR. Ranasinghe (The University of Melbourne, Australia) A Novel MAC Protocol for Power Efficient Short-Range Wireless Networking T. Y. Chui and W. G. Scanlon - University of Ulster, Northern Ireland ROC: A Wireless MAC Protocol for Solving the Moving Terminal Problem C.-H. Yeh (Queens University, Canada)

177

187

197

Part 8 Protocol Design and Mobility Support Variable-Radius Routing Protocols for High Throughput, Low Power, and Small Latency in Ad-hoc Wireless Networks C.-H. Yeh (Queens University, Canada)

215

XIV

An Overview and Comparative Evaluation of Wireless Protocols A. Mercier (Ecole Centrale d'Electronique, France) P. Minet (INRIA, France) L. George and Gilles Mercier (University of Paris, France)

228

Supporting Mobility in Distributed Proxy Server Architecture H. Lee, K. Chung (Kwangwoon University, Korea) and K. Kim (LG Electronics, Korea)

241

Part 9 Interoperability and Co-existence LAPP Enhancement Protocol X. Jin and J. Li (Xidian University, China) The Hybrid of Listen-Before-Talk and Adaptive Frequency Hopping for Coexistence of Bluetooth and IEEE 802.11 WLAN Y. Kim, B. Zhen and K. Jang (Samsung Advanced Institute of Technology, Korea) A Hybrid Architecture of UMTS and Bluetooth for Indoor Wireless/Mobile Communications T. Kwon, R. Kapoor, Y. Lee, M. Gerla (University of California , USA) and A. Zanella (Universita degli Studi de Padova, Italy)

253

263

273

Part 10 Multicarrier Systems Adaptive Link Adaptation for Multicarrier Systems F. Tang, S. Thoen, M. Engels andL. Deneire (Interuniversity Microelectronics Center, Belgium) On the Performance of the Ubiquitous Antennas for the Reception of COFDM Signals S. Okamura, S. Komaki (Osaka University, Japan) and M. Okada (Nara Institute of Science and Technology, Japan)

285

295

XV

Improved Automatic Frequency Control for OFDM H. Hosseini and B. Rohani (Genista Research, Singapore)

305

Part 11 Antennas and Interference Control A MIMO Architecture for Wireless Indoor Applications L. Giangaspero, G. Paltenghi (CEFRIEL, Italy) and Luigi Agarossi (Philips Research Monza, Italy)

317

Inter-cell Interference Effect on OFDM-based Wireless LAN P. Mahonen and A. Jamin (University of Oulu, Finland)

327

Part 12 Mobile Ad-Hoc Networks Analyzing Capacity Improvements in Wireless Networks by Relaying H. Karl and S. Mengesha (Technical University of Berlin, Germany)

339

Wireless LAN with Wireless Multihop Backbone Network (WMLAN) K. Mase, N. Karasawa, M. Kusumi, K. Nakano and M. Sengoku (Niigata University, Japan)

349

Author Index

359

Parti Mobile Computing

802.11: N O S T R I N G S A T T A C H E D

JOHN K FAWCETT Laboratory

for Communications Engineering, University of Trumpington Street, Cambridge CB2 1PZ, UK E-mail: [email protected]

Cambridge,

WILLIAM R SOWERBUTTS AT&T

Laboratories Cambridge Ltd, 24a Trumpington Cambridge CB1 1QA, UK E-mail: [email protected]

Street,

We describe the deployment strategy and application-level use of indoor and outdoor, point-to-point and point-to-multipoint 802.11b networks. For residential environments, applications combining radio networks with multimedia, location systems and home security systems are presented. We extend the home network paradigm with vehicular-based mobile computation where interference, shielding, range and power consumption issues are prevalent. The vehicle is considered both as a server—the user transports the 'master copy' of data—and as a client through disconnected operation and intelligent caching. Further, we discuss experiences of realising high bandwidth, low latency, fixed-broadband Internet access using outdoor 802.11b installations. Frequently neglected aspects of theoretic and practical network and data security are explored.

1

Introduction

Houses are for living, relaxing and spending time with the family. Increasingly these activities involve one or more personal computers. Parents often find themselves pruning the cabling sprawl laid down by their teenage offspring for head-to-head gaming. Interconnect technologies such as the Universal Serial Bus (USB) a , IrDA 6 , FireWire (IEEE 1394)6, X10 c , and Networked Surfaces7 seek to reduce cabling. Fully wireless mobility, eliminating short range line-ofsight requirements, is introduced by radio-based approaches including Bluetooth'*, PEN 1 (formerly Piconet), HomeRF e and the 802.11 4 family. This paper focuses on 802.11b and first discusses the outdoor installations used by the authors, comparing cost to the consumer, investment in infrastructure, performance and security to alternative always-on consumer broadband a

U S B Implementers Forum: http://www.usb.org/ infrared communication, http://www.irda.org/ c h o m e automation, http://www.xlO.org/ d short range radio, http://www.bluetooth.org/ e http://www.homerf.org/

3

4

provisions. Identical hardware finds further use within the home in Section 3 transporting bandwidth intensive video streams and delay-sensitive control signals. Applications and services enabled by mobile computing and wireless communications in a vehicle context are explored in Section 4. 2

Fixed Broadband Wireless

Employers seek to maximise staff productivity in order to lower operating overheads and remain competitive. Providing access from home to their corporate e-mail and filespace encourages employees to continue their work at home and allows them to remain in touch with colleagues. Broadband to the home can provide the rich information services expected of an office environment, allowing telecommuting and corporate expansion without the risk or financial burden of increasing office space. Access solutions making new use of existing wired networks—ISDN, leased lines, digital subscriber line (DSL) and cable modems—all provide sufficient bandwidth but incur repeated monthly costs from the service providers which own the infrastructure and provide connectivity. Offices already enjoy highbandwidth Internet connections and, from their rooftops, can offer line-ofsight radio connections directly to properties in the surrounding suburbs. Organisations providing employees with wireless connectivity to their networks can eliminate ISP subscription and most running costs in exchange for a small slice of the network administrator's time. Furthermore, broadband wireless 802.11b surpasses the bandwidth and latency offered by cable modems and DSL solutions (Table 1). The imminent release of 802.11a will further widen the price/performance gap. Most off-the-shelf 802.11b products do not offer external antenna connection jacks, which has severely limited the uptake of outdoor equipment and long-range deployments. Lucent Technologies' ORiNOCO product family' are one exception; SMC offer a similar product based on the Prism II chipset. The ORiNOCO cards cost approximately US$100 and outdoor antennae retail for US$120. The initial cost to the consumer is therefore comparable to that of cable modem and DSL services. The low latency of the wireless solution listed in Table 1 results from connections being direct rather than routed through a chain of ISP networks. The resulting responsive interactive behaviour of applications facilitates productive working, and was unanimously preferred by subjects in the trials. In the United Kingdom the European Telecommunications Standards In' h t t p : / / www.wavelan.com/

5 Table 1. Comparison of popular broadband solutions in Cambridge (UK) Technology ISDN Cable modem DSL 802.11b

Typical Bandwidth/kbps 64 / 128 512 512 4800

Latency/ms 90 40 50 2

Table 2. Outdoor 802.11b deployments in Cambridge Distance/km 0.25 0.5 1 2 4 7 10

Mode/Mbps

11/5.5 5.5/11

Latency/ms 2.0 2.0 2.0 2.0 2.0 2.1 2.2

stitute (ETSI) 3 and the Radio Authority (RA) h mandate that the effective isotropic radiated power (EIRP) output of systems in the 2.4 GHz band, measured at the antenna, be capped at 100 mW. This permits aerials of up to 12 dBi gain to be used with red* ORINOCO cards. The authors performed trials using Agere Systems 16 element vertically polarised directional Yagi antennae with 14 dBi gain, mast mounted at connection end points. Losses slightly exceeding 3 dBi were introduced by cabling, surge arrestors, band-pass niters and inline adaptors. Azimuthal gain plots 8 indicated the main beam width to be 7 degrees, with a half-power beam width of 30 degrees. Antennae were mounted on the lab roof and oriented to cover popular suburbs of Cambridge and near-by commuter villages. Similar antennae secured at domestic end-points were oriented back towards the lab. The established links are summarised in Table 2. Lucent quote a maximum achievable range of 25 km in ideal conditions with symmetric Yagi antennae if, as is often the case with domestic connections, a single megabit of bandwidth is tolerable. A line-of-sight to accommodate the Presnel bulge proved essential for links 9

http://www.etsi.org/ http://www.radio.gov.uk/ •"Red" ORiNOCO cards are designed for use with high-gain antennae and contain lower powered radio transmitters than their "black" counterparts, which are typically used by mobile nodes. h

6

exceeding 500 m; a partially obscured trajectory sufficed for shorter paths. Where geographic and topological constraints prevent direct connection, relay stations can be used to forward packets. Azimuthally orienting the antennae proved to be less troublesome than was expected; a good connection for a 10 km link was established simply by aiming at roughly the correct point on the horizon. Yagi antennae are polarised and sensitive to rotation about their lengthwise axis. The performance of the connections is substantially invariant to everyday weather conditions; link goodput does not suffer from rain fade, although the greater microwave opacity presented by fog causes a significant decrease in signal-to-noise ratio (SNR), ultimately resulting in drop-out. Thus far, each link is point-to-point. Trials of point-to-multipoint and mesh networks are underway, offering the potential for higher aggregate bandwidth and adaptive routing for fault and congestion resilience. In a dense mesh the links are generally shorter than those used above and omni-directional antennae are more appropriate, as illustrated in Figure 1. Mesh networks over 802.11b are not without compromise: the necessary intercommunication requires agreement over channel and encryption key. Spatial reuse of channels in non-interfering bands can increase total network capacity at the expense of potentially introducing nodes of articulation whose failure may partition the network. While parameter co-ordination is feasible within an organisation, the need for disjoint institutions to co-operate in synchrony may prove troublesome. The increased internal bandwidth is easily sufficient for voice-over-IP or video conferencing traffic. However, encrypted tunnels between the houses of participating individuals would be required if the internal bandwidth of the mesh is to be securely exploited. Link-layer wired equivalent privacy (WEP) offers a false sense of security as the integrity of the secret key is prone to compromise 10 . Point-to-point IPsec or Virtual Private Network (VPN) tunnels would provide end-to-end encryption but the human overhead of establishing per-tunnel shared keys scales poorly. Opportunistic Encryption^ solves this problem by dynamically generating shared secrets for IPsec connections as they are constructed. When one node needs a secure connection to another— for example when a voice-over-IP call is made—each fetches the other's public key from a trusted repository. Responses from the repository server are signed under a key whose public component is well-known. The signature is verified to prevent man in the middle attacks, and a shared secret is constructed by both nodes using the local private key and the peer's public key. This ma3

implemented, for example, in PreeSWAN. http://www.freeswan.org/

7

aa

Isn Figure 1. Mesh networking with 802.11b

terial is knowable only by the two communicating endpoints and suitable for deriving IPsec key material. The contributions to wide-area networking are the highly cost-effective provision of significant, reliable, always-on connectivity to the home. Wireless Internet Service Providers (ISP) are undercutting their wire-based counterparts in certain US cities although the current scale of operations leaves many neighbourhoods without coverage. The UK consumer has yet to be offered wireless broadband. Value for money and unobtrusive installation are the primary requirements of most homeowners; each can be satisfied by 802.11b hardware. Operating costs are minimal and employers can affordably provide subscription-free or subsidised bandwidth to employees. 3

Home Area Networking

PCs are becoming increasingly common in the domestic environment; they are used not only in their traditional office roles, but as entertainment centres and information gateways. The convenience and value of networking PCs in the home makes for familiar reading. However wired networking in the home is not as convenient as in an office as homes typically do not have either suitable cabling or cable trunking installed. This problem is particularly marked in the UK, where new-build houses

8

are rare and cables must be retro-fitted to existing properties. Hollow walls are uncommon, and the expense of buying and cabling network access points is overwhelmed by the cost of "making good" aesthetic aspects. Wireless technologies have a clear advantage in not requiring any structural modifications, especially in listed buildings where approval must be sought. The radio zone around an 802.11b node is sufficiently large to encompass an entire house, which allows the use of IBSS mode ad-hoc networks. Wireless networks make possible many applications which would otherwise be prohibitively expensive. The authors conducted comparison trials using a group of technical people with prior home Ethernet deployments. Hardware wireless access points and NAT bridges were eschewed in favour of Linux boxes, which offered greater configurability. These allowed the use of per-flow rate throttling, packet filtering and more secure firewalling, as well as routing onto the wide-area networks discussed in Section 2. The radio channel used for the WAN was chosen to avoid interference with the 802.11b network within the house. The trials revealed that popular applications of wired networks remained so with their wireless counterparts; file-sharing, gaming and Internet access were used extensively by all participants. An unexpectedly popular application was streaming media—several participants streamed live video encoded from a frame grabber and TV tuner to watch TV on their wireless laptops, in one case using a data projector and HiFi as an immersive home cinema. One participant reported installing a new digital TV aerial in his loft using the wireless video stream to guide alignment. Spooling the encoded video to disk allowed the system to be used as a video recorder for purposes of time shifting programming. Streaming music from storage to playback devices was also popular; one participant digitised his entire music collection of several hundred CDs, losslessly compressed, onto a 360 Gb RAID-5 array. This application has wide appeal as it makes one's music collection available anywhere in the house without the need to hunt for, or change, media. High-quality webcams connected to nearby embedded computers allow wireless video conferencing, and also double up as security cameras. Using the wireless wide-area network the information from these cameras could be easily transported off site. The privacy concerns of distributing live video from inside the house are addressed by using an asymmetric cipher to encrypt each frame on the capturing PC. The private component of the encrypting key is held securely and only released to the relevant authorities (eg, the police or insurance company) when the owner chooses to do so. Thus, even if the PC which encrypts the video is compromised, the spying intruder cannot decrypt previously streamed video data.

9 The wireless network also proved useful for data acquisition around the house. One house has curtains which open or close under computer control, and a wireless weather monitoring system from Oregon Scientific. Data acquisition cards in the embedded PCs throughout the house sample signals from microswitches concealed in each door and window, connected by short wires. Information from these sensors is fed over the wireless network for storage in a database and used for security applications—for example, the occupant is alerted upon leaving the house if any windows remain open. A second house uses microswitches and wall switches connected to small microcontroller boards which communicate over a shared RS485 serial bus. The wall switches become event generators controlling digitally dimmable lighting connected to the wireless network, replacing the traditional hardwired light switches in each room. However, the serial bus must still be wired to nearest the embedded 802.11 wireless node. A future study will replace this serial bus with a Bluetooth network. Bluetooth was chosen because its low power consumption allows battery powered switches which require infrequent cell replacement. Light switches could be glued to the wall since there is nolonger a requirement for any wires. These short-range Bluetooth networks would be bridged by the nearest embedded PC onto the 802.11 backbone network. The embedded PCs also poll local Active Badge 9 networks; an Active Badge is a battery-powered, wearable computing device capable of bidirectional infrared (IR) communication with fixed 'stations'. Badges beacon periodically to advertise their presence; their location is determined by which station can hear them. Occupants who choose to wear an Active Badge make their location available to the house computers, and can issue events by click buttons on the Badge. The house computers control Sentient Computing 5 applications based on the location and actions of occupants. For example, curtains, lighting, music and so on can all be automatically controlled. The Badges provide a two-button control set which users carry continuously. However, to support less cumbersome interaction with complex devices an 802.1 lb-enabled palmtop is used as a ubiquitous remote control and portable media player. 4

Car Area Networking

We have extended home and corporate networks into the private vehicles of several volunteers. Car-PCs provide navigation and guidance aids for the driver 2 and entertainment, information and workstation facilities to any passengers. The hardware hosts analog and digital I/O logic, creating a mobile

10

Figure 2. Car-PC and environmental sensors

data acquisition and information processing platform. Radio networking provides a convenient, high bandwidth transfer mechanism for use between the vehicle computer and a fixed base station, or amongst a collection of vehicles. Car-PCs axe operable both when driving and with the ignition off. 4J

Environmental Monitoring

Our Car-PCs (see Figure 2) continually log meteorological data and ambient atmospheric pollutant concentrations: every second the vehicle GPS is polled and measurements are taken from environmental sensors which measure ambient light, temperature, barometric pressure and humidity, as well as atmospheric levels of nitrogen dioxide, sulphur dioxide, carbon monoxide and ozone. Extensible Markup Language (XML) events are generated and serialised to a logfile. Some data have immediate application, for example if, while driving, a high Nitrogen Dioxide level is encountered the climate control unit can be directed to recirculate air and thus block out carcinogenic traffic fumes. The previous setting can be restored when pollutant concentrations return to safe levels. On arrival at home or work, a Car-PC uses the bandwidth afforded by the radio network to upload the logfiles accumulated since its previous synchronisation. The data are then useful to city planners striving to reduce urban pollution for students and tourists. Wired networks such as Ethernet and USB

11

would require trailing cables out to the parking lot; using removeable media such as compact flash or floppy disks would necessitate additional equipment in the car and still require human intervention to transfer the data. The range of Bluetooth is insufficient to communicate with our lab on the fourth floor and infra-red transducers are saturated by sunlight during daylight hours precluding either technology from application in this context. The convenience and automation of the 802.11b package is allowing city councils to consider installing the equipment in busses and taxis to collect data in a larger survey. 4-2

Entertainment

The convenience of the wireless network enables spontaneous downloading of music to the Car-PC system. The hassle-free user experience does not present a psychological barrier or otherwise deter use of the system—from their home or office PC the user has only to drag and drop audio files onto a icon depicting the car. An unduly tedious set-up procedure would be excessively cumbersome and cause users to elect to live with their existing configuration and avoid the agitation associated with makes changes. Vehicular use of 802.11b is not limited to file copying. Car-PCs operate an Active Badge station 9 which is used to identify passengers, distinguish theft and use by the owner and control the entertainment and navigation software. As I approach my driveway and the radio network connection with the computers in my house is restored real-time events can be passed to configure the house for my arrival. For example, the car stereo can communicate with a PC in the house to create the effect of the audio stream 'following-me' into the house. In order to achieve this the radio layer must recognise its peer and complete any handshaking in a timely fashion. For example, DECT supports appropriate ranges but is not designed for situations where recovery from prolonged signal loss is expected and consequently compliant devices may not notice the presence of another node for up to 20 minutes after communication becomes possible. Bluetooth and PEN devices remain alert to changing connectivity and could be used in place of 802.11b for this application. However, PEN has insufficient bandwidth to stream reasonable quality audio to the house, as is necessary if the track being played is not present in the user's home music collection. 4-3

Document Transfer

A car with a Car-PC can be used to transfer files, sneaker-net style, as the owner drives about. Using the car as a briefcase to transfer electronic documents simplifies to a drag-and-drop copy operation. More often, users save

12

and edit documents directly over the network using NFS or Samba services hosted by the Car-PC. The former modus operandi leaves the 'master copy' at work—a copy is taken away—suiting group-working but necessitating disconnected operation and update-merging techniques. The latter paradigm corresponds to single-owner documents and greatly simplifies the user's task of transporting files; consider, for example, a sales executive storing all his PDF brochures, client notes and his diary on a Car-PC.

4-4

Future Applications

A future car could exploit available broadband Internet connectivity by connecting to its manufacturer's website for code updates, safety advisories or simply to book a service or MOT. The computer might utilise location-based services, for example offering a search of near-by restaurants, service stations or digital map vendors to the occupants. Many such interactions can be triggered automatically, for example on approaching a city, a vehicle's computer might procure an enumeration of digital map distributors so the driver can easily purchase a detailed town plan before entering the city centre.

5

Conclusion

Outdoor 802.11b deployments are cost-effective and satisfy the bandwidth and latency requirements of telecommuters residing up to 10 km from their employer's office. Encrypting tunneling protocols safeguard personal and corporate data and scale linearly to mesh networks through opportunistic encryption. Within the home we enabled novel applications and improved existing ones through mobile and embedded computers, and reduced wiring clutter without compromising the decor or aesthetics. In a wireless house careful consideration must be given to the privacy of the occupants and the security of the network. For motor vehicles, navigation, sensing, data transport and entertainment applications based on Car-PCs and wireless communication were presented. The use of identical hardware at work, at home, between the two and in the car results in reduced costs. User interface familiarity additionally benefits employers by reducing the cognitive load of performing simple configuration, thus increasing productivity. Next-generation wireless standards are emerging, offering increased bandwidth and strengthened security.

13

Acknowledgements John Fawcett receives an EPSRC grant and a CASE award from AT&T Laboratories, Cambridge and is working on a PhD at the Laboratory for Communications Engineering. Thanks are due to Martin Brown for suppling some of the equipment and to Ian Wassell and Alastair Beresford for their assistance with radio regulatory compliance issues. References 1. F. Bennett, D. Clark, J.B. Evans, A. Hopper, A. Jones, and D. Leask. Piconet - embedded mobile networking. In IEEE Personal Communications, Vol. 4, No. 5, 8-15, 1997. 2. J.K. Fawcett and A.H. Jones. The road to driveable computing. Royal Institute of Navigators Journal, To appear. 3. J.K. Fawcett and W.R. Sowerbutts. Wireless network security. In press, 2001. 4. V. Hayes. IEEE standard for information technology - telecommunications and information exchange between systems - local and metropolitan networks - specific requirements - part 11: Wireless Ian medium access control (mac) and physical layer (phy) specifications: Higher speed physical layer (phy) extension in the 2.4 ghz band. Technical Report Designation 802.11b-1999, IEEE, 1999. 5. A. Hopper. Sentient computing. In The Royal Society Clifford Patterson Lecture, 1999. 6. G.A. Marazas. IEEE standard for a high performance serial bus. Technical Report Designation 1394-1995, IEEE, 1995. 7. J. Scott, F. Hoffmann, G. Mapp, M. Addlesee, and A. Hopper. Networked surfaces: A new concept in mobile networking. In 3rd IEEE Workshop on Mobile Computing Systems and Applications, 2000. 8. Lucent Technologies. User's guide for outdoor router outdoor antennas rev. c. In ftp://ftp. orinocowireless. com/pub/'docs/Wave A CCESS/MANUAL/ OR/oaig.c.pdf, 2000. 9. R. Want, A. Hopper, V. Falcao, and J. Gibbons. The active badge location system. ACM transations on Information Systems, 10(1):99-102, January 1992. 10. S. Fluhrer, I. Mantin and A. Shamir. Weaknesses in the Key Scheduling Algorithm of RC4. Selected Areas in Cryptography, 2001.

CLASSROOM IN THE ERA OF UBIQUITOUS COMPUTING SMART CLASSROOM

CHANGHAO JIANG, YUANCHUN SHI, GUANGYOU XU AND WEIKAIXIE Institute of Human Computer Interaction and Media Integration Computer Science Department, Tsinghua University Beijing, 100084, P.R.China E-mail: jiangch @media.cs. tsinghua.edu.en, shiyc @ tsinghua. edu. en, [email protected],

[email protected]

Abstract: This paper first presents four essential characteristics of futuristic classroom in the upcoming era of ubiquitous computing: natural user interface, automatic capture of class events and experience, context-awareness and proactive service, collaborative work support. Then it elaborates the details in the design and implementation of the ongoing Smart Classroom project. Finally, it concludes by some self-evaluation of the project's present accomplishment and description of its future research directions. Keywords: Ubiquitous Computing, Intelligent Environment, Multimodal Human- Computer Interaction, Smart Classroom

1

Introduction: From UBICOMP to Smart Classroom

Desktop and laptop have been the center of human-computer interaction since the late of last century. As is a typical situation of human's dialogue with computer that a single user sits in front of a screen with keyboard and pointing device, interacting with a collection of applications [Winograd 1999]. In this model, people often feel that the cumbersome lifeless box is only approachable through complex jargon that has nothing to do with the tasks for which they actually use computers. Too much of their attention is distracted from the real job to the box. Deeper contemplation on valuable matured technologies tells us: the most profound technologies are those that disappear, which means they weave themselves into the fabric of everyday life until they are indistinguishable from it [Weiser 1991]. We use them everyday, everywhere even without notice of them. Based on this point of view, computer is far from becoming part of our life. Mark Weiser first initiated the notion of Ubiquitous Computing (UBICOMP) at Xerox PARC [Weiser 1993], which envisioned, in the upcoming future, ubiquitous interconnected computing devices could be accessed everywhere and used

14

15

effortlessly, unobtrusively even without people's notice of them, just as electricity or telephones of today. This inspiring view of prospect has been accepted and spread so fast and widely that in a short time of a few years, many ambitious projects have been proposed and carried on to welcome the advent of UBICOMP. There are a bunch of branch research fields under the banner of UBICOMP, such as Mobile Computing, Wearable Computing and also Intelligent Environment, etc. The focus of this paper, Smart Classroom, belongs to the field of Intelligent Environment. But what is Intelligent Environment? In our point of view, we define it as an augmented spacious environment populated with many sensors, actuators and computing devices. These components are interwoven and integrated into a distributed computing system which is able to perceive its context through sensors, to execute intelligent logic on computing devices and serve its occupants by actuators. (In some research projects, Intelligent Environment is also referred as Interactive Space, Smart Space etc.) In researches of Intelligent Environment, there are several relevant and challenging issues need to be solved, such as the interconnection of computing devices on many different scales, the handling of various mobility problems caused by user's movement, and network protocol, software infrastructure, application substrates, user interfaces issues etc. Although many projects have been conducted in the name of Intelligent Environment, they have different emphases. Some focus on the integration of different sensing modalities [Coen 1999, HAL 2000], some aim at the adaptability of Intelligent Environment to user's preference [Mozer 1999], some are interested in automatic capture of events and rich interactions that occurs in an Intelligent Environment [eClass 2000, Adowd 2000], and some target at facilitating the collaboration of multi-user multi-device within a technology-rich environment [Interactive Workspace 2000, Fox 2000]. We can easily enumerate several other ongoing Intelligent Environment projects with different specializations, such as Georgia Tech's Aware Home [Aware Home 1999, Kidd 1999], IBM Research's DreamSpace [DreamSpace], Microsoft Research's EasyLiving [EasyLiving] etc. Our institute developed special interest in exploring the impact of ubiquitous computing to education. This leads to the project of Smart Classroom. The Smart Classroom is a physical experimental environment, which integrates multimodal human computer interface with CSCW modules collaborating through inter-agent communication language to provide a smart space for lecturer's natural use of computer to give class to distance learning students. In the rest of this paper, we're going to first present our views of futuristic classroom in UBICOMP. And then toward the ideal model of classroom, which sounds a little Utopian, we'll explain the idea and focus of our exploration. Later some details in the design and implementation of our present work will be illustrated. We'll conclude by a short description of our future goals.

16

2

What Should Classroom in The Era of UBICOMP Be Like?

Michael H. Coen from MIT Artificial Intelligence Lab said, "Predicting the future is notoriously difficult". [Coen 1999] Yes, we're not able to prescribe what future would be, but we're able to create toward what we think it would be. In our point of view, the following features are essential to a smart classroom in the era of UBICOMP, and will serve as the guidelines in our ongoing Smart Classroom project. We have generalized four characteristics of futuristic classroom, which are: natural user interface, automatic capture of class events and experience, context-awareness and proactive service, collaborative work support. 2.1

Natural user interface

As Mark Weiser has observed, "Applications are of course the whole point of ubiquitous computing". In accordance with this essence of UBICOMP, it is necessary for a smart classroom to free its occupant's attention to computer. To rescue people's energy from irrelevant interaction widi computer to the intentioned goal, allowing user's interaction with computer as naturally as possible is vital. In such a new paradigm of human-computer interaction, people input information into computer in their most familiar and accustomed ways like voice, gesture, eye-gaze, expressions etc. Auxiliary input devices like keyboard, mouse, are not necessary. In the reverse side, computer tends to serve people like an intelligent assistant. It utilizes technologies like projector display, voice synthesis, avatar, etc. This is what we call natural user interface. To get a clearer image, suppose a lecturer in the Smart Classroom conducting the class by voice. "Let's go to chapter two". Computer recognizes phonetic command and projects the wanted courseware of chapter two on display. Lecturer also uses hand gesture as a virtual mouse to annotate on the projected electronic board. Through combination of eye gaze (or finger pointing) and voice command, lecturer can zoomed in the image of an area in the projector to give emphasized explanation on a specific topic. 2.2

Automatic capture of class events and experience

This is what eClass project of Gatech called "automated capture, integration and access" problem. We use computer in classroom not only to improve die quality of teaching activity, but also to augment its capability, which was impracticable traditionally. The automatic capture of class event and experience belongs to such capabilities. It's not just record of video and audio in the environment, which is common in traditional distance learning-television broadcasting program. It

17

includes the record of group collaboration, multimedia events, multiple channels of human computer interaction, etc, all the events and experience that happened in the environment. The captured events and experience should be assembled into a kind of multimedia compound document. People can recreate the class experience by play the recorded multimedia compound document, and also can search a specific event or query knowledge within the compound document. This technology provides lecture content to students who are unable to attend the class in person, as well as to those who wish to review the materials later. For example, suppose a lecturer giving a class on Artificial Intelligence in a Smart Classroom. All the audio, video information, lecturer's annotation events, student's question events, Smart Classroom's controlling of lights, slides, etc, are recorded into a multimedia compound document. When a student wants to review the knowledge of Alpha-Beta Pruning algorithm, he can just query about it through his laptop computer, and rewind to the previous talk on it for a quick review and then comes back. After the class, students can also replay the document to recreate the class experience. 2.3

Context-awareness and proactive service

What is context-awareness? According to Dey & Abowd (from Gatech 1999), "context is any information that can be used to characterize the situation of an entity, where an entity can be a person, place, physical or computational object", "context-awareness is to perceive the context by system so as to provide task-relevant information and/or services to a user, wherever they may be". Which means the Intelligent Environment can understand user's intention not only based on audio-visual inputs, but also based on its situational information. Proactive service means to serve the user without his request. Proactive service is based on the Intelligent Environment's capability of Context-awareness. This model of service is disparate from traditional human-computer interaction paradigm, in which computer respond to human's explicit command. In the Intelligent Environment, the computer remembers the past, recognize the present, and predicate the future. It reasons human's intention through analysis of all the information from accumulated knowledge base. Then it tries to serve its occupants proactively with the reasoned intention of its user. For example, the lecturer is explaining a formula displayed on the electronic board. When the lecturer points at it and starts to talk about it, the computer understands that the lecturer is going to attract students' attention to the specific area of display. Then it zooms in the area containing that formula on the display without the need of lecturer's commanding "Zoom in this region". Another example, when the lecturer wants to have a student named Wang to give his opinion on a topic, he points at the student and says, "Wang, would you please say

18 something about what you think of this problem?" The computer then automatically focuses the video camera and microphone array on Wang and filters out the noise emitted from other spaces. 2.4

Collaborative work support

Class is essentially a collaborative procedure evolving multiple participants. In the environment of ubiquitous computing, the introduction of many interconnected computing devices and wide area network support enables us to extend beyond the space boundaries imposed by traditional classrooms. With this technological advance, collaborations of multi-user and multi-device can be possible. And the support for collaboration is becoming a requisite of a smart classroom. The collaborative work support of a Smart Classroom can be categorized into two classes. One is the collaboration of multiple attendants within the Smart Classroom holding various computing devices like, pen-based devices, hand-held devices and wearable computer etc. The other is the collaboration of remote participants and local attendants. The demand for collaboration support is so obvious that many commonly observed tasks in a classroom, such as group discussion, evolve the collaboration among multiple persons. There are some projects specialized in enabling and exploiting smart classroom's collaboration support, such as collaborative note-taking in both Gatech's eClass [eClass 2000] and Stanford's Interactive Workspaces [Interactive Workspaces 2000 ].

3

The Focus of Smart Classroom

Smart Classroom is a big project, every above-mentioned aspect of it is challenging and a long-term effort. Our institute has been committing itself to the research on multimodal human computer interfaces, CSCW in wide area network, and also multimedia integration. Based on our existing research results we have been investigating Smart Classroom's following features: natural user interface, automatic capture of classroom events and experience, and collaborative work support. So in initial phase of our project, we focus on applying our previous research achievements to realize an experimental environment. We have set up the physical experimental environment to demonstrate our idea and focus. In this Smart Classroom, we mainly aim at the following features: conducting lessons by means of gesture and voice command, capturing class events and operation such as manipulation on courseware, video-audio streams of class, etc, admission control of students using all kinds of mobile computing devices. In order to give clearer image

19

of our research, an elaboration of the physical experimental environment layout and its user-experience scenario are given as following. 3.1

The layout of Smart Classroom

Our Smart Classroom is physically built in a separate room of our lab. Several video cameras, microphones are installed in it to sense human's gesture, motion and utterance. According to UBICOMP's characteristic of invisibility, we deliberately removed all the computers out of sight. Two wall-sized projector displays are mounted on two vertically crossed walls. According to their purposes, they are called "Media Board" and "Student Board" separately. The Media Board is used for lecturer's use as a blackboard, on which prepared electronic courseware and lecturers' annotation are displayed. The Student Board is used for displaying the status and information of remote students, who are part of the class via Internet. The classroom is divided into two areas, complying with the real world classroom's model. One is the teaching area, where is close to the two boards and usually dominated by lecturer. The other is the audience area, where is the place for local students. Why are both remote students and local students supported in this room? The answer is simple, that we're complying with the philosophy of Natural and Augmented. Natural means we'll obey real-world model of classroom as much as possible to provide lecturer and students the feeling of reality and familiarity, which leads to the existence of local students. Augmented means we'll try to extend beyond the limitation imposed by the incapability of traditional technology, which is the reason for remote student.

AStudent Boards

^f

DDDDDD DDDDDD DDDDDD DDDDDD DDDDDD DDDDDD Audience Area

»"Teaching Area"!

Figure 1. Layout of Smart Classroom

In Smart Classroom, users' rights to use the room are mapped to their identification. There is an audio-visual identification module for identifying the users in this room and authorizing control right to lecturer. With the help of visual

20

motion-tracking module, Smart Classroom can be aware of its occupants' places in the room. Once a user identified as lecturer entering the teaching area, he is authorized to control the Smart Classroom by voice and gesture command. Lecturer can use hand-gesture as a virtual mouse on the Media Board to annotate, or add/move objects on the electronic board. He can also command linking in courseware, perform operations like scrolling pages, removing objects, granting speech right, etc by voice. All the lecturer's operation on the courseware and audio-video information captured in the Smart Classroom are automatically recorded and integrated to a multimedia compound document. The recorded information is simultaneously broadcasted to remote Students via Internet synchronously. Through software's application layer trans-coding and adaptive reliable multicast transport, remote students are promised to join the class with devices varied in computational power and display resolution through heterogeneous network varied in quality of service.

Figure 2. Typical scenario in Smart Classroom

3.2

A typical user experience scenario

The following is a typical user-experience scenario happened within the Smart Classroom. Multiple persons enter the room through the door. At the door, there is an audio-visual identification module identifying the entering person's identity through facial and voice identification. If the person is identified as lecturer, he is granted the control right of the Smart Classroom. The visual motion-track module tracks the lecturer's motion in the room. Once he steps into the teaching area, he will be able to use gesture and voice command to exploit the Smart Classroom to give lessons. Persons in the Smart Classroom other than lecturer are deemed as local students. When the lecturer is in the teaching area, he can start the class by just

21

saying, "Now let's start our class." The Smart Classroom then launches necessary modules such as Virtual Mouse agent, Same View agent (which will be talked about later). Lecturer loads prepared electronic courseware by utterance like, "Go to Chapter 1 of Multimedia course". The HTML-based courseware is then projected on the wall display. Lecturer can use hand-motion to stimulate the Virtual Mouse agent to annotate on the electronic board. Several type of hand gestures are assigned corresponding semantic meanings, which cause several operations like highlighting, annotating, adding pictures, remove object, executing links, scrolling pages etc, on the electronic board. Lecturer can also grant speech right to remote students by finger pointing or voice command, like "Li, please give us your opinion". On the Student Board, remote students' photos and some information as name, role, speech right etc are displayed. When a remote student requests for floor, his icon on the Student board twinkles. Once the lecturer grants the floor to a specific remote student, his video and audio streams are synchronously played both in the Smart Classroom and on other remote students' computers. 4

Details of Smart Classroom's Design and Implementation

The Smart Classroom is essentially a distributed parallel computing environment, in which many distributed software/hardware modules collaborate to accomplish specific jobs. Software infrastructure is the enabling technology to provide facilities for software components' collaboration. There are some alternative solutions to software infrastructure, such as Distributed Component-Oriented Model, like EJB, CORBA, DCOM, etc, and Multi-Agent Systems (MAS). In the context of Intelligent Environment, Multi-Agent System is more competent than Distributed Component-Oriented Model due to the following reasons: higher encapsulation level, faster evolution from design to implementation, easier development and debugging, and most importantly, more accordant to the need of dynamic reconfiguration and loose-coupling. 4.1

4.1 Software platform-OAA

In current stage, instead of developing our own Multi-Agent System, we choose to use SRI's famous open source MAS product, Open Agent Architecture (OAA) [OAA]. There're already many successful multimodal human-computer interaction projects built on OAA. Its delegating computing model also fits well in our need of software infrastructure. In OAA's delegating computing model, the network of distributed software modules is conceptualized as a dynamic community of agents, where multiple agents contribute services to the community. When external services or information are required by a given agent, instead of calling a

22

known subroutine or asking a specific agent to perform a task, the agent submits a high-level expression describing the needs and attributes of the request to a specialized Facilitator agent. The Facilitator agent will make decisions about which agents are available and capable of handling sub-parts of the request, and will manage all agent interactions required to handle the complex query. Such a distributed agent architecture allows the construction of systems that are more flexible and adaptable than distributed object frameworks. Individual agents can be dynamically added to the community, extending the functionality that the agent community can provide as a whole. The agent system is also able to adapt to available resources in a way that hard-coded distributed objects systems can't. 4.2

Five dedicated agents in Smart Classroom

In the schematic figure of our Smart Classroom there're five dedicated agents (except for the Facilitator agent of OAA). The Facial-voice identification agent is in charge of the Smart Classroom's login identification and authentication. When a person entering the room, he is required to place his face into a specific zone of a video camera's capture range, and speak a login word. The vision-part of the agent identifies the person by searching in a pre-trained user library, and the voice-part authenticates the identified person by voice-based speaker recognition. The motion-tracking agent is a computer vision-based agent. There's a pan-tilt video camera mounted on the upper side of the front wall, monitoring the whole range of the room. The motion-tracking agent receives video stream input from the camera and tracks the lecturer's position and movements in the room. When lecturer enters /leaves the teaching area, motion-tracking agent will signal the corresponding events to the agent society.

Voice Command Agent

^ilz-Vi

p)

"'' ' • Motion Track

Agent

Virtual Mouse Agent

Figure 3. Five dedicated agents in OAA model

23

The voice command support of Smart Classroom is realized by a speech recognition agent, which can perform speaker independent and continuous voice recognition. We use IBM's simplified Chinese version of ViaVoice SDK to wrap the voice recognition agent. The agent receives digitized signals from a wireless microphone, which is carried by the lecturer. And then recognizes its command within a dynamically loaded vocabulary set. Once a recognizable command is reached, the voice recognition agent dispatches that command to the agent society. Other corresponding agent is responsible for the execution of that command. The Virtual Mouse agent is used for handling hand-gesture, which stimulates the mouse events and shortcut command to activate operations on the playing courseware. It's also a vision-based agent. There are two video cameras specialized for virtual mouse event recognition. One is installed on top of the screen, the other is mounted on the ceiling of the room. Through detecting and analyzing 3D movements of hand, gestures can be recognized. The virtual mouse agent then dispatches the recognized mouse event or shortcut command to the agent society.

Figure 4. How virtual mouse agent works

SameView agent plays a core role in our Smart Classroom's pedagogical scenario. It is based on a legacy desktop application, namely SameView [Pei 1999, Liao 2000, Tan 2000], which is developed by media group of our institute. The purpose of SameView is a software for supporting multimedia based group discussion whose members are spaciously distributed and connected by heterogeneous networks. SameView has the following features: a shared MediaBoard (multimedia extensions to traditional electronic whiteboard), adaptive multimedia contents trans-coding according to terminal's network Qos and computing power, adaptive reliable multicast in wide range of heterogeneous

24

networks, live capture of video/audio streams and multimedia events into self-defined multimedia compound document, post-edit and playback of the captured multimedia compound document, self-equipped authoring tools for courseware-editing. In our Smart Classroom, we recur to the SameView's desktop version code as much as possible. We only revised some of its input/output user interface, such as adding a separate Student Board for display of remote students information and status by exploiting dual display adapter card support of Microsoft Windows 98/2000, projecting the Media Board in full-screen mode to remove the vestiges of desktop software with Windows-style menu, toolbar, title bar, etc. The most crucial reformation to Same View is the wrapping of it as an autonomous agent in Smart Classroom's agent community, which enables it to receive user's natural input from other dedicated agents like voice command recognition agent and virtual mouse agent, and then behave interactively. 5

Conclusion: Future Goals for Smart Classroom

Our current stage Smart Classroom is a primitive prototype of futuristic classrooms, which attempts to embody some of its distinguishing features like natural user interface, capture of class events and collaborative support. It is still far from a real Smart Classroom. Its resolutions to some key problems in Intelligent Environment are simple, intuitive and somewhat application- specific. Although many research issues need to be addressed in order to realize genuine Smart Classroom, we stride forward the first step toward the ambitious goal. In the near future, we'll make efforts in enhancing the Smart Classroom in some of the following aspects. Add more modalities and applications. We'll try to equip some more modalities of human inputs like vision-based tracker, embedded microphone array and various distributed sensors to sense human's context. And progress in the sensing technologies needs to be matched by progress in applications that use sensed information. Application is one of the key driving forces of technical advance. We'll conceive more realistic and useful scenarios in the Smart Classroom and also cooperate with different research groups whose application projects have high potential to take advantage of the capability of Smart Classroom. We believe the rise in the amount and sort of applications will enable generalization of Smart Classroom's design and implementation. Add a brain. The current design implementation of the Smart Classroom focus on the human's natural input to the computing environment, the next step is to move to a higher level and to give it the ability to understand. It is not just to utilize multimodal interface, but also to add-on context-aware intelligence. The classroom should be able to reason human's intention through analysis of all the gathered

25

inputs and proactively serve its occupants. There is some research studies in reasoning human's intention based on predefined grammars [Johnston 1998] or probabilistic statistical model. Each of them has innate weakness. We'll try to explore the combination of them. Add multi-user interaction. In the current stage Smart Classroom, there is only one user (lecturer) naturally interacting with the Intelligent Environment Other attendants are just observers or listeners and are not able to exploit the fascinating features of natural interaction. Because a class is bound to have multiple participants, to make a qualified Smart Classroom, we need to enhance the classroom's support for multi-user interaction. In our next step, we'll empower the classroom with capability to track and identify more than one user dynamically, and enable Smart Classroom's in-place service to every user in the room. Reference 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.

[Winograd 1999].Winograd, Terry. Toward a Human-Centered Interaction Architeture. 1999. http://www.graphics.stanford.edu/ projects/iwork/papers/ humcent/index.html [Weiser 1991]. Weiser, Mark. The Compu- ter for the 21st Century. Scientific American, pp. 94-10, September 1991. http://www.ubiq.com/hypertext/ weiser/Sci AmDraft3 .html [Weiser 1993]. Weiser, Mark. Ubiquitous Computing. IEEE Computer "Hot Topics", October 1993. http://www.ubiq.com/hypertext/weiser/ UbiCompHotTopics.html [Weiser 1994]. Mark Weiser. The world is not a desktop. Interactions, pages 7-8, January 1994. http://www.ubiq.com/hypertext/weiser/ ACMInteractions2.html [Coen 1999]. Coen, Michael. The Future Of Human-Computer Interaction or How I learned to stop worrying and love My Intelligent Room. IEEE Intelligent Systems. March/April. 1999. [HAL 2000] MIT AI Lab HAL Project (previous Intelligent Room project), 2000. http://www.ai.mit.edu/projects/hal [Mozer 1999]. Mozer, Michael C. An intelligent environment must be adaptive. IEEE Intelligent Systems. Mar/Apr. 1999 [eClass 2000] Georgia Tech, eClass Project (previous Classroom 2000) 2000 http://www.cc.gatech.edu/fce/eClass/ [Abowd 2000]. Abowd, Gregory. Classroom 2000: An experiment with the instrumentation of a living educational environment. IBM Systms Journal, Vol. 38. No.4 [Interactive Workspaces 2000]. Stanford Interactive Workspaces Project. http://graphics.stanford.edu/projects/iwork

26

11. 12. 13.

14. 15. 16. 17.

18. 19. 20. 21.

[Fox 2000] Fox, Armando, et al. Integrating Information Appliances into an Interactive Workspace, IEEE CG&A, May/June 2000 [Aware Home 1999]. Georgia Tech, Aware Home Project,1999 http://www.cc.gatech. edu/fce/house [Kidd 1999]. Kidd, Cory D., Robert J. Orr, Gregory D. Abowd, et al. The Aware Home: A Living Laboratory for Ubiquitous Computing Research. In the Proceedings of the Second International Workshop on Cooperative Buildings CoBuild'99. Position paper, October 1999. [DreamSpace] IBM Research. http://www.research.ibm.com/natural/ dreamspace/index.html [EasyLiving] Microsoft Research, http://www.research.microsoft.com/vision [OAA]. SRI. http://www.ai.sri.com/~oaa [Johnston 1998]. Johnston, Michael. Unification-based multimodal parsing. In the Proceedings of the 17th International Conference on Computational Linguistics and the 36th Annual Meeting of the Association for Computational Linguistics (COLING-ACL 98), August 98, ACL Press, 624-630. [Ren 2000]. Ren, Haibing, et al. Spatio-temporal appearance modeling and recognition of continuous dynamic hand gestures. Chinese Journal of Computers (in Chinese), 1999. Vol 23, No. 8, Agu.2000, pp.824-828. [Pei 1999]. Pei, Yunzhang, Liu, Yan, Shi,Yuanchun, Xu, Guangyou. Totally Ordered Reliable Multicast for Whiteboard Application. In proceedings of the 4th International Workshop on CSCW in Design, Paris, France, 1999. [Tan 2000]. Tan, Kun, Shi, Yuanchun, Xu, Guangyou. A practical semantic reliable multicast architecture. In proceedings of the third international conference on multimodal interfaces,BeiJing,China,2000 [Liao 2000]. Liao, Chunyuan, Shi, Yuanchun, Xu, Guangyou. AMTM - An Adaptive Multimedia Transport Model. In proceeding of SPIE International Symposia on Voice, Video and Data Communication, Boston, Nov. 5-8, 2000.

H Y B R I D JINI FOR LIMITED DEVICES

VINCENT LENDERS, POLLY HUANG AND MEN MUHEIM E-mail:

{lenders,

ETH Zurich huang}@tik.ee.ethz.ch,

[email protected]

We envision a future of heterogeneous mobile devices collaborating spontaneously and bringing convenience to life. Taking a practical approach, we study the feasibility of integrating limited devices into the Jini, Java, and Linux paradigm. With careful evaluation and system hacking, we manage to fit the software stack, excluding Java's Remote Method Invocation (RMI) support, in a small 1.85-Mbyte space. However, we also identify that including RMI support will exhaust all the remaining space. We propose, as a short-term solution, to re-implement Jini with light-weight communication alternatives and to use a hybrid lookup server to inter-operate RMI- and non-RMI-supported devices. For the long term, we advocate minimalism and call for community-wide standardization effort to the system software development.

1

Introduction

We envision a future that computers and electronics will be mobile and collaborate spontaneously, bringing convenience to life. Two immediate problems before realizing such a future are ways to handle heterogeneity and mobility of these devices. These devices, for example mini-laptops, mobile phones, and home appliances have very distinct hardware and software profiles. The task of configuring and inter-operating them will be too difficult for average users, not to mention that these devices are likely to be mobile and whenever they enter a new environment they need to be re-configured. To counter the problems of automatic configuration and seamless interoperation, the industry and research community have offered a few solutions 1>2 that encompass (1) a resource discovery/management protocol and (2) a framework of uniform programming interface and code base. One prominent example is the Jini 3 and Java 4 combination. In the Jini/Java paradigm, devices can come and discover (or be discovered by) resource in the network automatically. While responding, the devices register proxies allowing other members in the network to operate them. As a feasibility study, we are building such a Jini/Java-based spontaneous network with various types of devices, mainly full-fledged computers, mobile PCs, and limited electronics. Connecting limited electronics is the most interesting and challenging case. By limited electronics, we mean personal, home, or office electronics. They are essentially micro-computers with compact sys27

28

tem architectures and limited flash memory, and sometimes also referred to as embedded systems. Taking a hands-on approach, we set ourselves off to install the Jini/Java stack on a development board for embedded systems. This turns out to be a non-trivial task and gives rise to a serious system software problem - its size. It is profoundly difficult just to fit a native Jini/Java/Linux stack onto a typical embedded system board, not to mention extra applications required to run on top of the system stack. After intensive system hacking and careful evaluation, we find bottleneck of the size problem being Java's Remote Method Invocation (RMI) support. It is humongous, relatively to other parts, and current implementation of Jini depends on it. While RMI offers elaborated remote execution support, it is not really necessary for communication in Jini which can be rather primitive. We propose, as a solution to scale system software size, a light-weight Jini implementation based on IP socket or XML. That is to implement Jini without using any RMI interfaces and avoid including RMI support in the software stack at all. To warrant backward compatibility with existing implementation, we take a hybrid approach to differentiate existing RMI-based and the lightweight implementations, thus enabling execution of non-RMI proxies on RMIsupported devices but not the other way around. More importantly, we call for standardization activities towards defining minimum system software stack for mobile devices. In the meantime, we urge developers to refrain from RMI and the community to re-examine its scalability. In short, our contribution includes a) a successful port of system kernel and free-license VM for limited devices, b) a qualitative and quantitative evaluation of a minimum software stack, c) a 35% reduction of the Jini/Java/Linux stack without RMI support to 1.85 Mbytes, and d) a light-weight hybrid Jini solution to overcome the system software size problem. 2

Approach

Jini devices require hardware platforms with network connectivity and the ability to execute Java code. Two different minimal hardware approaches are possible. The first solution is an embedded device with network connectivity that runs a Java Virtual Machine (VM) on top of any kind of operating system. Typical embedded devices of this kind are the Developer Board LX from Axis 5 and the NetSC520 Demonstration Platform from AMD 6 . Both platforms run embedded Linux and communicate via Ethernet. The second solution is to use embedded boards with Java processors. Unlike traditional microprocessors,

29

which must convert Java byte-code into the processor's native language, these processors can operate directly from Java byte-code. aJile Systems 7 provides a development board with their Java processor called aJile. Another solution is provided by Systronix 8 . They developed a hardware Java platform with their Tiny Network Module (TINI). We decided to use the Developer Board LX from Axis for our further implementations. This board provides the best scalability regarding resources, size, and price. It has the following features: 100MIPS 32 bit CPU, Ethernet 10/100Mbps, 2 RS-232 serial ports, 2 parallel ports, 2 Mbyte FLASH and 8 Mbyte DRAM. The FLASH size can be updated to 4 Mbytes. Operating Jini on this platform with restricted resources requires strict size limitations regarding the operating system and the VM. Linux suits well as target operating system for our device. The two main reasons are the compact size of Linux and its open license policy. We use a Linux port as operating system. Different Java versions from different parties are candidates to be ported to the developer board. Sun Microsystems introduces the Java 2 Micro Edition (J2ME) 9 as Java platform for limited devices. J2ME is divided in two complementary configurations: the Connected Limited Device Configuration (CLDC) and the Connected Device Configuration (CDC). Both configurations introduce a new VM. The CDC contains the C virtual machine (CVM) and the CLDC contains the K virtual machine (KVM). CLDC would probably better suit for the developer board. Kaffe 10 is an open implementation of Java for embedded and desktop systems. It is an implementation of the PersonalJava 3.0 specification and requires no source code licenses from Sun Microsystems. The major advantages of Kaffe are: compact size, easiness of porting to new platforms (operating systems and architecture), and it is open source software. For these reasons, we decided to use Kaffe instead of J2ME for our implementation. 3

Evaluation

The main task of this work is to fit the required software on the 4-Mbyte FLASH ROM. The required software includes the operating system, VM, Jini core classes and Jini service classes. Embedded Linux. The CPU on the developer board is the ETRAX 100 LX. The source code of Linux 2.4 includes the CRIS architecture for ETRAX 100 LX. Axis provides a small distribution of Linux for their developer board. We have optimized this distribution and reduced the size of the operating system (kernel and basic utilities) to 800 Kbytes (55 % of the original size).

30

OS and Utils

Kaffe VM Klasses.jar

Jini and Service

Figure 1. Flash Reservation (4 Mbytes)

Breakdown of component sizes in the FLASH ROM is depicted in Figure 1. Kaffe V M . As described in the last section, we use Kaffe as Java VM. Kaffe has not been ported to the CRIS architecture yet and it is part of our contribution to port Kaffe to this new architecture. The classes in Kaffe are bundled in eight jar files. The core Java classes are part of the Klasses.jar file. The original classes are shown in Table 1. For memory saving reasons, we port only a subset of the core Java classes. The awt, applet, beans, math, and sql packages are removed since they are not useful for the given board as a Jini device. Table 1 shows the size of the different packages and the removed packages. The original size of the Java core classes in Kaffe was 886 Kbytes. We could reduce the size of the core classes to 525 Kbytes. The VM was compiled with the gcc-cris cross-compiler, gcc-cris is a port of the GCC 11 to the CRIS architecture. We had to compile with static li-

31 package

package

size [K]

jarcompressed size [K]

lang awt util net security io text applet beans math sql

428 944 588 168 184 324 112 20 104 20 80

97 213 133 38 42 73 25 5 24 5 18

lang awt util net security io text applet beans jar management

76 60 100 104 40 420 40 36 40 12 12

17 14 23 24 9 95 9 8 9 3 3

3920

886

removed

Java r

r r r r

kaffe

Total

r

r r r

525

Table 1. Package Size and Reconfiguration of Kaffe

braries since there is no support for shared libraries at the current time. The executable Kaffe VM for the CRIS architecture takes about 400 Kbytes of memory space. Figure 1 shows the size of the Kaffe VM and the Java classes in respect to the whole FLASH ROM. Jini. Jini is just a set of additional Java classes. The basic Jini class files are packed in three jar files: jini-core.jar, jini-ext.jar, and sun-util.jar. The Jini core classes are part of jini-core.jar and jini-ext.jar. The classes in sun-

32 Package net .jini.core, discovery net .jini.core.entry

Function support for unicast discovery definition of E n t r y objects for type definition support of remote events net.jini.core.event support classes t o implement net .jini.core.lease leasing net.j ini. cor e. lookup interface a n d support classes t o communicate t o t h e lookup service net .jini. core .transaction support classes for clients of t h e transaction manager net.jini.core, transaction .server support classes for services t h a t participate in transactions

RMI yes no yes yes yes

yes yes

Table 2. The Jini Technology Core Platform (JCP)

Package net .jini. admin net.j ini. discovery net.jini.entry net.jini.event net .jini. lease net.jini.lookup net.j ini. lookup. entry net.jini.space

Function interface to provide common ways of administrations support for unicast and multicast discovery utilities for Entry objects classes related to the event mailbox service renewal of leases assisting classes for lookup service joining and communication attribute types that services can use for registration with the lookup service JavaSpace support

RMI yes yes no yes yes yes no yes

Table 3. The Jini Technology Extended Platform (JXP)

util.jar are not part of the official Jini distribution. These classes are helper classes from Sun to help developers to write their services and clients. The major problem we encountered with the Jini core classes is that they require Remote Method Invocation (RMI). Our light-weight Java port does not support RMI. RMI requires too many computational resources to run on the board. In Table 2 and 3, the core packages of Jini are listed with their RMI dependence.

33

Our solution is to remove the RMI dependence from the Jini classes and to replace RMFs functionality with other mechanisms that scale for embedded devices and our Kaffe VM. Regarding the size restrictions, we have to fit the light-weight Jini and service classes in around 2.3 Mbytes of memory space. Next we discuss the solution in detail. 4

Solution

To solve the size problem, we can, on one extreme, add more memory on the limited devices or on the other extreme take the surrogate approach 12 that suggests to leave the limited devices as they are and to enable their participation through a Jini-capable delegate. Adding memory results in higher manufacturing cost which is not desirable from the industry's point of view. The surrogate approach, although solves the size problem, requires configuration of the Jini delegate on limited devices, which brings us back to the configuration problem that Jini is designed to eliminate at the first place; thus not desirable from the users' point of view. We propose a hybrid solution that solves the size, cost, and configuration problems. Having identified the bottleneck of the system stack, we suggest to replace communication part of Jini implementation, currently done by heavyweight Java RMI, with light-weight alternatives, such as IP socket or XML. In Table 2 and 3, we list the RMI-dependent subset that we intend to tackle. Changes to ensure minimum discovery and maintenance will influence both the limited devices and the lookup server. Thus, a hybrid lookup server, referred to as LUS+, needs to be in place to handle both native Jini and non-RMI Jini devices in the network. In case a non-RMI device joining a Jini network which is not capable of handling RMI-based and non-RMI-based discovery, one will have to fall back on the add-more-memory or surrogate approach. There is yet a backward compatibility problem. In case of a non-RMI device joining a Jini network assuming RMI fully supported, a LUS+ may naively pass RMI-based proxies to the non-RMI device and expect them to work. The solution will be to filter out RMI-based ones at the lookup server so limited devices will not receive proxies that they could not handle. Note that filtering of non-RMI proxies is not necessary. Devices with RMI support will not have any problem executing the non-RMI proxies. It is obvious that we cannot afford to communicate limited devices using space-hungry RMI. Although the proposed light-weight hybrid solution helps alleviate the problem, there remains several critical questions to address. For instance, do we absolutely need RMI as a means of communication in the

34

context of Jini or is RMI inevitably large in size? For the moment, we would urge developers to refrain from RMI, but for the long term, the community needs to take a close look at the scalability of RMI and to work towards defining a minimum set of communication interfaces so that limited devices (an important element in the future mobile network) can move and collaborate effortlessly. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.

Microsoft, Universal Plug and Play Forum, http://www.upnp.org Object Management Group, Inc., CORBA Forum, http://www.corba.org Community Resource for Jini Technology, http://www.jini.org Sun Microsystems, Source for Java Technology, http://java.sun.com Axis Communications, Developer Board LX, http://www.axis.com/products/dev_board/index.htm. AMD, NetSC520 Demonstration Platform, http://www.amd.com/products/epd/desiging/evalboards/21.netsc520e/index.html aJile Products, aJile, http://www.ajile.com/products.htm Systronix, TINI, http://www.systronix.com/home.htm Sun Microsystems, Java 2 Micro Edition, http://www.java.sun.com/j2me/, 2001. Transvirtual, Kaffe, http://www.kaffe.org, 2001. GNU, The GNU Compiler Collection (gee), http://www.gnu.org/directory/gcc.html, 2001.

1 2 . Jini Community, Surrogate Project, http://developer.jini.org/exchange/projects/surrogate/, 2001

Part 2 Quality of Service and Wireless Internet

A DIFFSERV-BASED CLASSIFICATION SCHEME FOR INTERNET TRAFFIC OVER WIRELESS LINKS DMITRIS SKYRIANOGLOU, NIKOS PASSAS AND STAVROULA KAMPOURIDOU University of Athens, Communication Networks Laboratory, TYPA Buildings, Panepistimioupolis, Athens 15784, Greece E-mail; (dimiski \ passas | s.kampouridou}® di.uoa.gr QoS provision for Internet traffic over a wireless link is a challenging issue. In order to cope with the upcoming difficulties, European 1ST project WINE (Wireless Internet NEtworks) introduced WAL (Wireless Adaptation Layer) as an intermediate layer that will adapt and fulfill the requirements of IP traffic over an underlying wireless system (HIPERLAN/2, Bluetooth, IEEE 802.11). WAL utilizes a set of modules such as ARQ, FEC, Header Compression, SNOOP, Fragmentation and QoS that can be set-up and combined together so as to improve the system performance and compensate for the poor transmission quality of the wireless link. The parameters of WAL modules can dynamically adapt to the changing wireless channel conditions so as to maintain an acceptable level of service. The operation of WAL modules is supervised by the WAL Coordinator. This work focuses on QoS provision within WAL and proposes a QoS classification scheme based on a DiffServ approach. According to this scheme, incoming traffic is classified using the DiffServ CodePoint (DSCP) of the IP Header. Several aspects of the proposed classification scheme are discussed and alternative approaches are proposed.

1

Introduction

During the last few years, Internet technology has evolved as the main way towards the new generation of telecommunication networks. The development of the version 6 of the Internet Protocol (IP) [1], that extends the functionality of the current version 4 (IPv4), provides new features and facilities such as expanded addressing capabilities, simpler header format, improved support for extensions and options, and a sophisticated interface for quality of service (QoS) provisioning. A challenging issue nowadays is how this QoS provisioning can be supported on an end-to-end basis. On the other hand the growing demand of users for mobility support has led to a considerable development of wireless communications. Especially wireless local area networks (WLANs) offering wireless indoor communication are becoming more and more popular and tend to replace the classic wired LANs. To this end, many standards from different organizations or fora have been specified (e.g., [2, 3, 4]) in order to extend the capabilities of wireless networks and to offer more advanced services. The evolution of Internet technology has also affected the development of WLANs considerably so one of the major issues today is the optimization of WLANs to support Internet services, especially for advanced IPv6 services.

37

38

The main problem for wireless communications is the unpredictable and errorprone behavior of wireless channels. Interference, fading, atmospheric conditions and terminal mobility may radically affect the quality of the channel. Under these circumstances the Transmission Control Protocol (TCP), the main Internet transport protocol used mainly for non-real-time applications, does not appear to be a suitable choice. Since TCP always assumes that packet losses are due to network congestion, it activates its congestion avoidance back-off mechanism, hence dramatically decreasing system throughput. In order to cope with this problem, numerous proposals have been made (e.g., [5, 6]). For real-time applications, on the other hand, where the User Datagram Protocol (UDP) is used, the wireless link should be able to meet the delay and loss requirements of these applications. The WINE (Wireless Internet NEtworks) is a project, sponsored by the European Commission, under the Information Society Technologies (1ST) programme, that deals with the performance enhancement of Internet protocols over WLANs. The core of the project's proposal is a "shim" layer, namely the Wireless Adaptation Layer (WAL) (Figure 1) [7]. 2

WAL Architecture

As its name implies, WAL adapts its behaviour according to the higher layers' requirements and observed channel conditions. It provides a uniform interface to IP, while being independent of the underlying wireless network technology. Within the project, three popular WLANs have been selected for study: Bluetooth [2], IEEE802.il [3], and HTPERLAN/2 [4]. From a functional point of view, WAL could be described as a performance enhancing proxy (PEP) [8] that improves the performance of Internet protocols operating over wireless shared access LANs. A novel feature of the WAL is an abstraction used for service provisioning at the link layer. Each IP datagram is classified into classes and associations. Service provision in the WAL is based on these two concepts. A WAL class defines the service offered to a particular type of application traffic (e.g., audio/video streaming, bulk transfer, interactive transfer, Web), and the sequence of link layer modules (protocol components) that provide such a service. The module list for every class is completely defined so that every WAL terminal uses the same module order within the same class. An association identifies a stream of IP datagrams belonging to the same class and destined to or originated from a specific mobile terminal (MT), i.e., WAL_Association = <WAL_Class, MT_Id>. A fair allocation of bandwidth can be easily achieved if based on a peruser operation. In addition, services for particular users can be customized to meet their QoS requirements and to implement a differentiated-charging policy. The WAL Coordinator shown in Figure 1 can be viewed as the central "intelligence" of the WAL. Both downstream and upstream traffic passes through the WAL Coordinator before being processed by other modules. In the downstream flow, the WAL Coordinator intercepts IP datagrams and decides on the sequence of

39

modules that these datagrams should pass through, as well as the parameters of these modules. In the upstream flow, the WAL Coordinator accepts WAL frames (encapsulated IP datagrams) and passes them through the sequence of modules, associated with, the class, in the reverse order. Information about the modules' sequence as well as the required module parameters is contained in the WAL header, used for communication between peer WAL entities.

Data Exchange: (UpStrearfi

(DotvnStmarr} (j) v

W irel e ss D riv e r I nte rfac e (Generic Ethernet interface)

Figure 1. WAL. Architecture

The QoS module provides flow isolation and fairness guarantees.through traffic shaping and scheduling. Traffic regulation before scheduling is essential to assure that a traffic stream is not affected by unpredictable bursts of packets injected by

40 other sources. Therefore, it is included in the QoS module design so as not to rely on QoS performed in other layers. Modules X/Y/Z comprise a pool of modules, aiming to improve performance in a number of ways. This pool includes error control modules such as Forward Error Correction (FEC) and Automatic Repeat reQuest (ARQ), a Snoop module [6] for TCP performance improvement, Header Compression module, and a Fragmentation module to reduce packet loss probability. Other modules may be included, in later versions of the WAL, to further improve the overall performance. Finally, in order to interface with a number of wireless drivers of different platforms, one LLCT (Logical Link Control Translator) module for each different platform has been introduced. The main functions of this module manage the connection status with the wireless driver, and ensure the stream conversions toward the wireless driver. 3

Classification of IP Traffic within WAL

Setting the class in each different association is one of the main and most importance functions of the WAL Coordinator. We refer to this function as "classification". The idea behind classification is that different IP data flows are classified into different classes according to their traffic characteristics and QoS demands. In this way WAL can provide each flow with a different treatment and a different level of service that will better meet its needs. Further to this, using classification WAL reduces significantly the amount of state information that needs to be maintained and facilitates the handling of many different data streams that can be grouped and handled as a single stream. 3.1

The proposed classification scheme

The classification scheme proposed here is based on the Differentiated Services (DS) architecture [9, 10]. According to the DS model, IP traffic entering a network is classified and possibly shaped and conditioned at the boundaries of the network, and assigned to different behaviour aggregates. Each behaviour aggregate is identified by a single 8-bit DS code point (DSCP), contained in the IP header. Within the core of the network, packets are forwarded according to the Per-Hop Behaviour (PHB) associated with the DSCP. The main advantage of DS architecture is that it achieves scalability, since state information for per-application flows or per-customer forwarding discipline does not need to be maintained within the core network. The WINE system can be viewed as a DS domain with the AP acting as the DS boundary node, interconnecting the WINE system with the core network or other DS domains. AP maintains a Service Level Agreement (SLA) with these neighbouring domains so that downlink (AP-to-MTs) traffic entering WINE system

41

is classified and treated appropriately. Furthermore, AP maintains a service contract with each MT within the WINE system. This contract determines how the uplink (MTs-to-AP) traffic is treated within WINE and how it is conditioned and marked with the corresponding DSCP before it is forwarded to the neighboring domain or the core network. So far two different PHB have been proposed within IETF, Assured Forwarding (AF) [11] and Expedited Forwarding (EF) [12]. The AF PHB provides delivery of IP packets in four, independently forwarded, AF classes. Within each AF class, an IP packet can be assigned one of three different levels of drop presence. So there are four classes with three different drop probabilities each. Therefore, twelve DSCPs have been reserved for this PHB group. Table 1. Recommended codepoints for the AF PHB (xx=currently unused).

Low Drop Medium Drop High Drop

Class 1 OOlOlOxx OOllOOxx OOlllOxx

Class 2 OlOOlOxx OlOlOOxx 0101lOxx

Class 3 OllOlOxx OlllOOxx OllllOxx

Class 4 lOOOlOxx lOOlOOxx 1001lOxx

The EF PHB can be used to build a low latency, assured bandwidth, end-to-end service through DS domains. To support this service, it is required in every transit node, that the aggregate's maximal arrival rate is less than that aggregate's minimal departure rate. This service appears to the endpoints like a point-to-point connection or a "virtual leased line" but it is not preferable for WAL since, due to its requirements, it is very demanding and may reserve a considerable part of the scarce available bandwidth of the wireless channel. Therefore, within WAL, IP packets with the EF PHB DSCP can be re-marked and mapped to AF PHB Class 1 All other DSCP not belonging to AF PHB group can be mapped to the default PHB (DSCP = 000000XX) which corresponds to the standard Best-Effort Service. The proposed approach for the WINE system makes use of the AF PHB in order to classify the IP flows and provide a different service to each traffic class. In this approach each type of application is mapped to one of the four main classes of AF PHB. We identify 4 types of applications: •

Conversational (e.g., Voice over IP, Tele-conference)

•

Streaming (e.g., Video/Audio streaming)

•

Interactive (e.g., Web Browsing)

•

Best effort (e.g., Mail, Telnet, FTP)

These types of applications are mapped on the AF/Default PHB DSCPs and classes as shown in Table 2

42 Table 2. Mapping of Application Types to AF/Default PHB DSCPs and Classes.

Application Type Conversational

Streaming

Interactive

Best effort

3.2

DS CodePoint (OOI)OIOXX (OOl)lOOXX (001)110XX (OIO)OIOXX (OIO)IOOXX (010)110XX (Oll)OlOXX (Oll)lOOXX (Oll)llOXX (100)010XX (100)100XX (100)110XX OOOOOOXX

Class

AFI AFII AFITI AFIV Default PHB

WAL Module chains

According to the DSCP mapping, presented above, each IP packet entering the WAL is classified to the corresponding PHB Class, which in turn corresponds to a chain of WAL modules. [7]. After classification the WAL Coordinator forwards the incoming packets to the first module in the chain. Each module in the chain performs its corresponding action on the packet (e.g., it adds CRC bytes for FEC), appends a module header and returns the packet to WAL Coordinator that passes it to the next module in the chain. At the receiving side the packet traverses the module chain in the reverse order. Within each module the packet is again processed and the module header is stripped. The modules of each chain are suitably chosen so as to facilitate the service provision required for each class. So far the available modules within WAL are: •

ARQ module (ARQ)

•

Fragmentation module (FRM)

•

IP Header Compression module (HC)

•

SNOOP module (SNP)

•

FEC module (FEC)

•

QoS module (always present and last in the module chain)

43

The proposed module chains for each class are shown in Figure 2. Note that for AF in and AF IV/Default Classes two alternative module chains are proposed. The second ones (those including SNOOP module) can be utilized only for TCP applications. TCP data flows can be recognized by the Protocol Type field of the IP header. Note also that the QoS module is always present and last in every module chain. QoS performs packet scheduling within WAL, hence it is put last in the chain so as to guarantee that the scheduled packets will be ready for transmission and will not experience any further delays within WAL due to processing or queuing in other modules.

AFI:

V HC

AFH:

HC

AFUI::

( HC )

FEC

QoS

HC

H FRM )

*i ARQ )

[ SNP

HC

FEC

AFIV/DefaultPHB:

I HC

QoS

AF rVTDcfault PHB TCP Applications:

\ SNP

HC

AFm TCP Applications:

QoS

QoS

*{ FEC j

*f QoS )

QoS HC: Header Compression FEC: Forward Error FRM: Fragmentation ARQ: Automatic Repeat Request SNP: SNOOP QoS: Quality of Service

Figure 2. Proposed module chains for each Class

According to the DS architecture each module chain implements the PHB for each class. The choice of the modules in each chain is based on the nature and requirements of each application type, such as delay and loss requirements. The order of the modules is chosen so that the preceding modules do not cause conflict or degrade the operation of the following ones. These module chains are just a few of the possible chains that may be utilized in order to improve the performance of WAL. Since the classification scheme can be further refined by making use of the subclasses of the AF PHB classes, a set of possible module chains that will better fit to the specific requirements of each (sub)class can be defined.

44 3.3

Association Establishment and Maintenance

For each new WAL Association a new instance of the module chain corresponding to the class of the association is created. The parameters of each module (e.g., ARQ window size, FEC CRC code, etc.) in the chain are chosen so as to provide the association with appropriate level of service, taking into account current channel conditions. This is an additional feature of the proposed architecture compared with the standard DS architecture, where all IP packets of each behavior aggregate get exactly the same forwarding PHB within a DS domain. Since the wireless channel conditions may vary in time for each MT, the parameters of the modules within a chain may be dynamically adapted to the new channel conditions so as to maintain the agreed level of service of each association. Packet arrives from upper layer

Look for packet destination IP address Discard packet

-f&Mr Registered MTs

Yes

No Discard packetOR check for new class

Yes Get modules list & parameters for this class

Access class entry

Establish association wih theMT

Add new association

o

defisit ons

Send packet using this

Figure 3. Action performed when an IP packet arrives from upper layer

Jmc Established associations

45

In order to keep track of the established associations and their characteristics, WAL maintains within, the Management Information Base (MIB), information concerning each new association. As shown in Figure 3, for downlink traffic, each time a new IP packet arrives at the WAL, WAL checks if the destination MT is registered. If yes, WAL determines the service class of the packet and checks whether the MT supports this service class and whether there already exists an association for this class. If a new association has to be established, WAL Coordinator determines the module chain and the module parameters corresponding to the class of service and the current channel conditions. Then it informs the MT about the new association. After the successful establishment of the association the packet is forwarded for transmission through the new association. 4

Conclusions - Future Work

A DS-based classification approach was presented. The proposed scheme tries to incorporate and extend the features of DS architecture within the WINE system. According to the proposed scheme, each IP flow is classified to a different class of service. For each class a different WAL module chain is instantiated. The parameters of the modules are dynamically adapted to reflect the changes of the wireless channel conditions. The presented scheme is sufficiently flexible and adaptive and can be further refined so as to support more sophisticated classification schemes. The future plans of this work include simulation study of the classification scheme, performance comparison with other existing schemes and the refinement of the proposed classification mechanism. Acknowledgements This work has been performed in the framework of the project 1ST-1999-10028 "Wireless Internet NEtworks (WINE)", which is funded by the European Community. The authors would like to acknowledge the contributions of their colleagues from VTT Electronics, University of Oulu, Philips Research Monza, University of Rome "La Sapienza", AQL, Cefriel, Intracom S.A., University of Athens, Acorde, University of Cantabria, and The Queen's University of Belfast. References 1. S. Deering and R. Hinden, "Internet Protocol, Version 6 (IPv6) Specification," RFC 2460, December 1998. 2. Bluetooth consortium, "Specification of the Bluetooth system," vl.OB, Dec. 1999.

46 3. "IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications," P802.ll, Nov. 1997. 4. ETSI TR 101 683, "Broadband Radio Access Networks (BRAN); HIPERLAN Type 2, System Overview," February 2000. 5. A. Bakre and B.R. Badrinath, "I-TCP: Indirect TCP for Mobile Hosts," in Proc. ofICDCS95, May 1995. 6. H. Balakrishnan, S. Seshan, and R.H. Katz, "Improving Reliable Transport and Handoff Performance in Cellular Wireless Networks," Wireless Networks, vol. 1, no. 4, pp. 469-481, December 1995. 7. P.Mahonen, N.Passas, et al., "Platform-Independent IP Transmission over Wireless Networks: The WINE Approach", accepted for publication in the IEEE Personal Communications Magazine. 8. J. Border, M. Kojo, J. Griner, G. Montenegro, and Z. Shelby, "Performance Enhancing Proxies," Internet Draft (work in progress), , October 2000. 9. RFC2474, K. Nichols, S. Blake, F. Baker, D. Black, "Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers", December 1998. 10. RFC2475, S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, W. Weiss. "An Architecture for Differentiated Service", December 1998. 11. RFC2597, J. Heinanen, F. Baker, W. Weiss, J. Wroclawski, "Assured Forwarding PHB Group", June 1999. 12. RFC2598, V. Jacobson, K. Nichols, K. Poduri, "An Expedited Forwarding PHB", June 1999.

A MANAGEMENT ENTITY FOR IMPROVING SERVICE QUALITY IN MOBILE AD-HOC NETWORKS MARC BECHLER, BERNHARD HURLER, VERENA KAHMANN, LARS WOLF Institute of Telematics, University of Karlsruhe Zirkel 2, 76128 Karlsruhe, Germany Phone: +49 721 608 [6397 I 6397 I 8072 I 8104] - Fax: +49 721 388097 Email: [bechler I hurler I kahmann I [email protected] Quality of Service in mobile ad-hoc networks has been considered so far with respect to improving routing protocols mostly. However, routing table updates may often be timeconsuming and re-routing does not work in the case a mobile device offering an service (e.g. an Internet gateway) is shutting down. In this paper, we propose a flexible Management Entity that integrates both network state changes and service changes into a management information base within the IP layer. Service discovery protocols are helpful in detecting service information. We show that our Management Entity allows switching network technologies efficiently and thus can avoid breaking TCP connections.

1

Introduction

Due to recent advances in hardware technology and the evolution in mobile and wireless networking technologies, mobile devices gain more and more importance. Cellular phones, PDAs, or laptops become more and more powerful and popular as they are nowadays able to perform user's everyday tasks. On the networking point of view, there was a growing demand to interconnect those devices using ad-hoc networking technology, which allows a simplified use of the mobile devices without any configuration effort. Currently, several networking technologies for setting up ad-hoc networks coexist, such as Bluetooth, IEEE 802.11 (in ad-hoc mode), IrDA, or HIPERLAN. With UMTS, new and more powerful communication technologies might come up in the future, resulting in an extremely heterogeneous and variegated environment. A third trend can be seen in the growing importance of protocols for service location, especially for ad-hoc networks. Since static or manual configuration of servers, proxies or gateways may be impossible or at least annoying in many cases, those services have to be discovered within the ad-hoc networking range. Several service discovery protocols have been developed to alleviate the need of configuration for the user, like Jini [1], Bluetooth Service Discovery Protocol (SDP) [2], and the Service Location Protocol (SLP) [3] of the IETF, for example. In current communication systems those functionalities coexist in terms of independent mechanisms and protocols. However, a combination and interaction of those mechanisms and protocols can provide for synergistic effects and, thus, significantly improve the quality of service support in an ad-hoc network. In this paper, we propose a flexible and efficient Management Entity (ME) located within mobile end systems operating in heterogeneous ad-hoc networking environments. The basis 47

48

of this entity is a multiplexing mechanism that enables a mobile device to hand over connections between different bearers in a seamless way. A Management Information Base (MIB) supporting the multiplexing mechanism holds out information on the current status of each attached network interface card. In order to improve the service quality in ad-hoc networks, the ME exploits service location information, provided by an arbitrary service location protocol. This functionality, which is integrated within the operating system's network stack, works transparent to the applications so that they need not to be modified. Additionally, legacy applications are also supported and can take full advantage of this handover mechanism. 2

Approaches to Improve the Perceived Quality of Service

In contrast to [4], which shows the need for switching between different networks for wireless infrastructure-based technologies, multi-hop ad-hoc networks meet several new challenges: They are characterized by a highly dynamic topology without any infrastructure-based communication technologies. Communication should be possible even if the networking topology changes rapidly, as mobile devices sending data are moving around as well as mobile routers and mobile devices receiving the data. Mobile devices should also be able to communicate with each other using different networking technologies, resulting in a highly heterogeneous networking environment. Figure 1 demonstrates this heterogeneity in a typical ad-hoc networking scenario with five mobile devices. They are partially interconnected using IEEE 802.11 (in ad-hoc mode) and Bluetooth technology. Additionally, two mobile devices - the cellular phone on the left, and laptop B on the right-hand side - are connected to the Internet: The cellular phone uses GSM communication technology (represented by link (2) to the base station, could be alternatively GPRS, UMTS, DECT, etc.), whereas laptop B has a high-speed ADSL communication link (1) to an Internet Service Provider ISP2. In such a scenario, an improvement of the endto-end quality of service in mobile and wireless ad-hoc networks can be achieved by exploiting the following two aspects: Utilization of Heterogeneity and consideration of service location information. 2.1

Utilization of Heterogeneous Environments

Although suitable routing protocols enable communication in multi-hop ad-hoc networks (e.g., DSR, TORA, AODV, c.f. [5] for an overview), communication paths between sender and receiver can break when the network is topology is partitioned due to the movements of the nodes. Consider the following situation in our scenario in Figure 1: In order to conserve energy, laptop A in the ad-hoc network initially communicates with laptop B using Bluetooth via the PDA (communication path (3)-(4)). If B becomes unreachable - for example when A moves out of the coverage of the Bluetooth network or when the PDA is switched off - communica-

49

tion is no longer possible. A's TCP connections will time out, even if B is still reachable using the IEEE 802.11 link (5) (which would increase the power consumption of both laptops A and B). The reason is that a TCP connection is uniquely identified by a quadruple (IP address A, port A, IP address B, port B) and switching to another network interface results in a new source IP address related with this interface. Internet.

A

U^L^tf?

802.11 fj (ad-hoc)y**

BT Figure 1. Ad-hoc Networking Scenario

However, it is also harmful to change the bindings of IP addresses to a networking device due to three reasons: • The mobile node becomes unreachable as the new address bound to the network interface might be topologically incorrect • The process of binding, unbinding, and the update of the internal routing table(s) is not very efficient. • Caching of ARP information is not possible as the matching between IP address and MAC address changes after the modification. Thus, a seamless and transparent switching mechanism between different networking interfaces is needed, which is currently not supported by common operating system's TCP/IP implementations. 2.2

Utilization of Service Information

Several service location protocols have been developed in the last few months. Among the most famous are Jini, UPnP (Universal Plug and Play), Bluetooth SDP,

50

Salutation and SLP. Most service location protocols may be used in ad-hoc networks (see [6] for a characterization of the protocols). Some protocols (Jini, UPnP) even provide service access in addition to service discovery, in case of Jini even without the need of pre-configured drivers for a service. For ad-hoc networks it is also important that a central service manager is not required because in a dynamic environment a centralized entity is always a single point of failure. Thus, service location protocols that implement distributed service managers or enabling direct discovery of services at a particular device should be preferred in ad-hoc networks. Our architecture does not depend on a particular service discovery protocol; however, some protocols are restricted on particular technologies (e.g. Bluetooth SDP). As described in the previous section, seamless switching between different networks is a basic feature for improving the quality of a perceived service. However, the heterogeneity also implies that the services are also distributed over the accessible ad-hoc networks. Due to device mobility, the services (in each network) need to be regularly discovered and their availability is not ensured as mobile devices (providing services) can be frequently switched off and on by their users. In the scenario described in figure 1, the laptop A has a connection to the Internet, using a gateway from ISP2. This gateway can be reached by means of IP routing via laptop B. If B is switched off (or becomes unreachable due to a link breakdown) A's connection to the Internet terminates. In this case, the service location protocol running on A has to discover an alternative proxy providing Internet access service in the heterogeneous ad-hoc network. In this case, a proxy from ISPl will be used which can be reached via the route (6)-(7)-(2). Afterwards, the network settings of A need to be reconfigured and the (communicating) application must be restarted as the source address might have changed. Of course, similar scenarios are conceivable with various other services, such as wireless and mobile telephony in ad-hoc networks using registration services. In our example, we only need to discover a service and retrieve information about it. This information is promoted to the IP layer via the MIB, so in our scenario, the IP layer will use the information that ISPl can be used as another proxy if ISP2 is unreachable. Thus, the ongoing communication with hosts in the Internet can be handed over seamlessly to the new proxy and need not be reestablished. 2.3

Architecture: The Management Entity

We designed and implemented a Management Entity (ME) that is able to avoid such breakdowns of ongoing connections as described in the example above by handing over the established communication links to a new transmission path. It is also able to consider service location information to find alternative services within the adhoc network. Figure 2 shows the integration and the components of this ME in more detail. On top of the communication layers are the applications that use the network protocols for communication. Additionally, an application daemon implementing a

51 service location protocol is running on the end system, which discovers the available services within the ad-hoc networks the mobile device is currently participating. Within common operating systems, such as Windows 2000 or Linux, TCP and/or UDP is used at the transport layer, and IP at the network layer, located above the link layer. The lowest layers, physical layer and link layer, realize the different transmission technologies for ad-hoc networks. As can be seen in Figure 2, our ME covers link layer, network layer, and aspects of the application layer. This results in an advantageous behavior of the communication protocols, as information from other layers could be utilized for optimizing the protocol mechanisms of the network layer, resulting in an improved quality of service support for the applications.

Application Layer

Service Location Protocol

Application Process l _

m Transport Layer

m "

Application Process

Application Process

; i

;>

.k

>

1 t

\:

TCP

UDP

TCP

k

—=,"""' Networking Layer

'

f

'k —

"

IP i

MIB

Link Layer/ Physical Layer

•


1 > Device-Multiplexer I . i

'

/\. . 1 ^""'Y^l»ifc_ ( ^ o m t o r ^ M 802.11 BI Bluetooth 1

. ' IrDA

Figure 2. The Management Entity

On the link layer, a Monitor collects relevant data from the different networks. As an example, the monitoring of an IEEE 802.11-based ad-hoc network comprises parameters like the availability, or the current packet loss rate (as a result of a calculation based on various other physical parameters such as signal strengths of different access points, etc.). Those parameters are stored in a Management Information Base (MIB), which is updated by the Monitor. Up to now, we only consider whether an attached networking interface is currently able to send and receive data to/from other mobile stations. A Service Location Protocol running in user space also interacts with the MIB. The MIB is regularly updated with the discovered services, i.e., the ME has an overview of the currently available services within the ad-hoc network and their location (in combination with the routing protocol, the ME knows which networking interface it has to use for communication with this service). If a

52

communication link breaks, the ME is able to find an alternative path to the receiver using another network - assuming that the receiver is accessible in general. If the mobile device uses a particular service within the network and the communication with the device that supports this service fails (e.g., when it is switched off), the ME has immediate access to the information about other stations offering a similar service in the ad-hoc network. The ME is then able to tunnel the packets to another device supporting this service. 2.3.1

Multiplexer

The basic element in our ME for achieving a seamless handover between different communication technologies is a Multiplexer, which is also shown in Figure 2. Note that the ME addresses end-to-end quality of service aspects. In multi-hop ad-hoc networks, communication paths between sender and receiver can change frequently due to the dynamics of the network topology. In this case, the ad-hoc routing protocol is responsible for ensuring the adaptation of the forwarding behavior from each node to the reconfiguration of the network. If the receiver becomes unreachable with the actually used networking technology (which can be detected by overflowing transmission buffers), the Multiplexer queries the MIB for alternative communication technologies supporting a communication path to the receiver. In order to achieve transparency for the applications, the TCP/UDP ports and the IP addresses that are in use must not change if the IP packets are sent via another network device, as described above. Hence, our Multiplexer encapsulates IP datagrams in new IP packets with the IP address of the chosen networking device as the source address. This tunneling mechanism, which is the main principle of Mobile IP, is performed by using a redirector within the multiplexing entity, that calls the IP stack a second time and, thus, redirects every packet to another networking device which sends them out. Due to the encapsulation of packets, the new created packet could exceed the maximum size for IP packets. Hence, the created IP packets must be additionally segmented and reassembled if necessary. When an encapsulated packet arrives at the receiving node, the Multiplexer first establishes a tunnel back to the sender (if such a tunnel does not yet exist), as we assume that a breakdown affects the direction from sender to receiver and vice versa. Afterwards, the Multiplexer unpacks the original IP packet and forwards it to the upper layer protocols using the original mechanisms of the IP implementation. Notice that our approach, all communicating mobile nodes must have a ME. 3

Related Work

Heterogeneous network environments are basically grasped as a challenge, not as a chance for QoS support. There are several projects focusing on the provision of an improved support for mobile hosts at different layers of the network stack, such as [7]. However, none of them addresses the end-to-end QoS aspect in mobile ad-

53

hoc networks. A different approach is to reconfigure the network dynamically, which is the basic of, e.g., [8]. Mobility support for computers is achieved by a dynamic network reconfiguration. The authors define an abstract set of device characteristics, which could also be useful for our Management Entity. In order to achieve mobility, guards within the operating system detect the validation and invalidation of the device characteristics. Depending on those parameters, the system enables an "intelligent adaptation" on network layer, transport layer, and application layer. At the network layer, the adaptation process reacts to invalidations of routes by modifying the routing tables within the kernel of the operating systems. This is different to our approach as we provide a more flexible approach by directly choosing the NIC without changing the bindings from IP addresses and routes. In contrast to our approach, QoS architectures for mobile ad-hoc networks try to support QoS by managing the resources within each node. For example, the INSIGNIA architecture, an in-band signaling architecture for QoS support in mobile ad-hoc networks [9], the resources within the mobile nodes along a communication path are reserved in a request/reserve manner before communication starts. Rerouting packets due to the mobility of nodes results in a flow restoration along the new path. Although those approaches support QoS for communication, they do not utilize the fact that a sending mobile node can choose between different transmission technologies if the mobile node is able to switch between them seamlessly. In the area of the IETF many drafts deal with the mobility support for IP and the challenges in MANETs. Most relevant for our work is the work in the MANET working group of the IETF, which basically tries to cope the heterogeneity with routing protocols for mobile ad-hoc networks. However, end-to-end quality of service aspects are currently not addressed in their work. The Intentional Naming System [10] developed by a research group at the MIT also allows to seamlessly handle node or service mobility. There, a late binding mechanism allows finding the appropriate service at runtime. An overlay network of resolvers uses this binding for routing. The emphasis in this approach lies on the naming system, whereas we do not change naming and therefore need not change applications. 4

Implementation and Evaluation

In order to evaluate our ME, we implemented a prototype using Linux. We used the Netfilter mechanism of the Linux networking architecture to receive, examine and re-inject IP packets. The implementation of the part that controls the ME is completely operating in userspace, which leads to a maximum of independence from future kernel reorganizations. To receive the packets with a userspace process we used the queue target of Netfilter. A so called acting part of the userspace process listens to the state of MIB, where it gathers important information about link quality. It initially decides to change to another networking device by setting up a tunnel over the new link and adding a new rule to the routing table, which sends all packets

54

to the communication partner over the new tunnel device. The so called reacting part listens for incoming tunneled packets (i.e., packets with 0x04 in the protocol field of the IP header). If such a packet arrives, the reacting part immediately creates a "reverse tunnel" and updates the routing table as well. All information necessary for this procedure can be extracted from incoming tunneled packets. Of course, the implementation in the userspace is very time sensitive. Therefore, we implemented very efficient functions to establish tunnels, to set up links, to add IP addresses and to update routing tables. In order to increase performance, it was important as well to infiltrate packets into the kernel as fast as possible to avoid buffer overflows. Although the implementation works in userspace, it is efficient enough to manage even high data rates. In order to evaluate our approach, we used two laptops, equipped with IEEE 802.11b PC Cards running in 2Mbit/s ad-hoc mode. Unfortunately, we had no Bluetooth PC Cards at our disposal, so we used a twisted pair cross-over cable instead. In our tests, we examined the impact of our Multiplexer on the behavior of UDP and TCP packet streams with different data rates. At the beginning, a stream was sent from one laptop to the other using the Ethernet link. After 4.170 seconds, we requested the multiplexing entity to tunnel the stream via the wireless link. In order to get exact performance measurements of the switching process to an alternative networking technology, we avoided a detection of overflowing buffers to identify a broken link. We also ensured the correctness by unplugging the cable. After 8.5 s, the stream was stopped. Figure 3 and 4 represent two basic aspects of our results. Figure 3 shows a spot of the behavior of a TCP stream with 1 Mbit/s and a packet size of 512 byte in the relevant time range of 4 s and 5 s (switching was initiated at 4170 ms). Firstly, it shows that the TCP connection does not break when the Ethernet cable is unplugged. Secondly, the switching to the wireless link has little impact on the TCP stream itself. Notice that we avoided TCP's slow start by using traffic shaping within the end system. After sending a TCP packet, the ACK returned before the next TCP packet is sent. Figure 4 shows our examinations of the delay that is generated by the Multiplexer in the interesting spot of 4.1 s and 4.3 s. We used a UDP data stream with 1 Mbit/s and a packet size of 512 byte. We measured the time when a packet arrived at the receiving laptop by tapping the processor time (i.e., the CPU cycles); thus, our measurements should be very precise. There is a only very short delay when the Multiplexer switches from the wireline to the wireless communication link. Of course, the measurements presented here are only a brief glance. Basically, the results show that the Multiplexer has hardly any impact on the packet jitter: Data rates with less than 512kbit/s (512 byte packet size) and packet sizes of more than 1408 byte result an no significant delay, as the tunnel is set up between the transmission of two packets. Hence, the measurements show that our Multiplexer is able to switch quickly between different network interfaces, which might be even faster than discovering new routes in ad-hoc networks after a network reconfiguration.

55

However, a much more significant delay occurs when a broken link should be detected, e.g., by overflowing transmission buffers.

CO

°- 1240 1190 1140 4000

4100

4200

4300

4400

4500

4600

4700

4800

4900

5000

4260

4280

4300

Time (ms) Figure 3. Behavior of a TCP Stream with 1 Mbit/s (512 byte Packetsize)

1180 4100

4120

4140

4160

4180

4200

4220

4240

Time (ms)

Figure 4. Performance of a 1 Mbit/s UDP Stream (512 byte Packetsize)

56 5

Conclusion

In this paper we presented a Management Entity for improving service quality in mobile ad-hoc networks. In contrast to other approaches such as QoS based routing, our approach is able to compensate link breakdowns by switching to alternative adhoc networking technologies in a seamless way - even if the communication partner is no further accessible via the current networking technology. In combination with information generated by a service location protocol, the QoS support can be significantly improved as the availability of services in ad-hoc networks also vary over time due to device mobility. Future work has to be done by defining common interfaces for our MIB, as well as to extend the MIB by various other parameters that allow a more exact prediction of the link characteristics. Also, we are currently improve our evaluation for the impact of service location protocols on the QoS support. Additionally, we have to consider upcoming communication technologies for ad-hoc networking (such as Bluetooth) as soon as they are available and affordable. References 1. Sun Microsystems: Jini Connection Technology, http://www.sun.com/jini, 2001 2. Bluetooth Consortium: Specification of the Bluetooth System Core Version 1.0 B: Part E, Service Discovery Protocol (SDP), November 1999 3. E. Guttman, C. Perkins, J. Veizades, and M. Day: Service Location Protocol, Version 2. RFC 2608, Internet Engineering Task Force, June 1999 4. M. Bechler, H. Ritter: A flexible Multiplexing Mechanism for Supporting Quality of Service in Mobile Environments. Proceedings of the 34 Annual Hawai'i International Conference on System Sciences (HICSS), Maui, January 2001 5. E. M. Royer, C. K. Toh: A Review of Current Routing Protocols for Ad Hoc Mobile Wireless Networks. IEEE Personal Communications, April 1999 6. C. Bettstetter and C. Renner: A Comparison of Service Discovery Protocols and Implementation of the Service Location Protocol. In Proceedings of EUNICE 2000, 6th Open European Summer School, Twente, September 2000 7. R. H. Katz, E. A. Brewer, E. Amir, H. Balakrishnan, A. Fox, S. Gribble, T. Hodes, D. Jiang, G. T. Nguyen, V. N. Padmanabhan, M. Stemm, "The Bay Area Research Wireless Access Network (BARWAN)", Proceedings ofCOMPCON Conference, April 1996 8. J. Inouye, J. Binkley, J. Walpole, "Dynamic Network Reconfiguration Support for Mobile Computers", Proceedings of the 3rd Annual ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom '97), Budapest, September 1997 9. S.-B. Lee, A. T. Campbell: INSIGNIA: In-Band Signaling Support for QoS in Mobile Ad Hoc Networks. Proceedings of the 5th International Workshop on Mobile Multimedia Communication (MoMUC'98), Berlin, October 1998 10. W. Adjie-Winoto, E. Schwartz, H. Balakrishnan, J. Lilley, "The Design and Implementation of an Intentional Naming System", in Operating Systems Review, 34(5): pp 186 — 201, December 1999

INTEGRATING IPV4 AND IPV6 IN WIRELESS NETWORKS EDGARD JAMHOUR Pontificia Universida.de Catolica do Parana, Curitiba, PR, Brazil E-mail: [email protected] This paper addresses the issue of implementing Mobile IP in wireless networks using private IP addresses. In the first part of the paper we review various concepts relating the use of private IP addresses in Mobile IP networks, including the limitations introduced by using NAT. In the second part, we propose a novel approach for implementing Mobile IP with private IPv4 addresses called "Transparent IPv6" (TIP6, for short). The TIP6 approach consists in virtually assigning IPv6 addresses to IPv4 hosts without modifying end user's or network devices. We show that the TIP6 approach extends standard NAT implementations by permitting a mobile host with private IP address to receive connections from other mobile or fixed host through the Internet. The TIP6 approach permits an IPv4 host to communicate with other IPv4 hosts in TIP6 networks or directly with "true" IPv6 hosts. This feature will permit the coexistence of emerging IPv6 Mobile IP implementations with existent IPv4 Networks.

1

Introduction

Wireless networks are supposed to extend the IP connectivity to portable devices, such as laptops, pocket computers and cellular phones. New standards, such as IEEE 802.11, solve the wireless problem connectivity at the layer 2 level. However, most applications still requires IP connectivity in order to communicate with fixed or wireless devices through the Internet. In this paper, we address the issue of implementing Mobile IP in wireless networks using private IP addresses [1,8]. This issue has become particularly important because of the shortage of IP version 4 (IPv4) addresses [7]. A possible solution for the IPv4 shortage problem is to implement wireless IP networks using private IP addresses, such as defined by RFC 1918 [6]. In order to permit mobile hosts to communicate through the Internet, techniques for implementing network address translation (NAT) are required [9]. However, NAT imposes limits to a mobile host operation, because there is no way to initiate a connection with it (refer to section 3). Because of the NAT limitations, the IP version 6 (IPv6) is considered as a concrete alternative to manage the IPv4 addressing problem [2,4]. IPv6 backbones are not absolutely required, because there are techniques that permit automatically creating tunnels between two IPv6 networks through an IPv4 network. That enables the immediate use of IPv6 addresses over the current Internet infrastructure. However, the IPv6 adoption would require changes in network

57

58

equipments and operating systems of wireless devices. To address this problem, this paper presents a novel approach for assigning IPv6 addresses to IPv4 devices without any hardware or software modification. This approach is called in this paper "Transparent IPv6" (TIP6, for short). The main idea is to provide all the benefits of IPv6 addressing without introducing changes in the existing Mobile IPv4 infrastructure. This paper is structured as follows. Sections 2 to 4 review fundamental concepts such as mobile IP standard, private IP addresses, NAT and IPv6. The TIP6 approach is presented in the section 5. 2

Mobile IP and Wireless Networks

This section reviews the Mobile IP standard and defines the terminology used in the remaining of this paper. Mobile IP standard [1,8] treats the problem that may arise when a host changes its IP address during a communication. A mobile host changes its IP address because the IP protocol [3] assumes that each IP network identifier is related to a specific physical network. Mobile hosts are connected to physical networks by RF links that attaches the mobile host to an "access point" device within a coverage area (see Figure 1). The IP protocol is implemented over RF exactly as it would be implemented in a wired network. However, when the mobile host changes its position, it can connect to another access point. When it happens, the mobile host can potentially be attached to a different physical network (depending on distance between the access points, i.e., when access points are interconnected by routers). If the mobile host connects to another physical network, it must change its IP. Mobile IP solves this problem by using a tunneling technique. In this approach, each mobile host has two IP addresses. One IP is related to its "home network" (where the mobile host is registered), and does not change when the host changes its position. The second IP is related to a "foreign network", and changes each time the host attaches to a different physical network. This second IP is called COA (Care-Of-Address). 802.11 corrpliant device

Home Agent

H3

Internet Wirod Network Backbone

ho

V;

Foreign Agent

Figure 1. WLANs inter-communication using Mobile IP.

The router attached to the mobile host at the foreign network is called "foreign agent". The router at the home network is called "home agent". The home agent is a

59

special router, responsible for authenticating the mobile host, and keeping an internal table mapping the COA to the home IP address of every mobile host. Mobile IP specifies that is up to the mobile host the responsibility of informing the home agent mat it has changed its COA. For doing this the mobile host send a "binding update" message to the home agent each time it changes its COA. The message is delivered to the home agent by the foreign agent. From a non-mobile host viewpoint, a mobile host is identified by its home IP address. Packets from the Internet are delivered to the mobile host through a tunnel that follows the hosts while it changes its position and attaches to different networks. Please refer to [1] for more details about the Mobile IP standard terminology and operation. 3

Private Addresses and NAT

Private IP addresses are defined by RFC 1918 [6]. Hosts with private IP addresses can exchange information with other hosts connected to the Internet "only" by using IP address translators such as Proxy or NAT (Network Address Translation) [9]. The operation principle for both devices is almost the same: they replace private IP addresses in the packets delivered to the Internet by public IP addresses. Operating this way, they permit to use the same public IP address for several devices (in general, thousands of private IP addresses per public IP). An important limitation, however, is that a host with a private IP can act only as a client, because its address is not visible by hosts outside the private network. A Proxy is typically implemented as an application server. That means that the proxy requires the client to redirect IP packets to it in order to operate (i.e., Proxy is not transparent to the client). NAT, by the other hand, is typically implemented on routers. The NAT device is usually seen as the default gateway for the hosts with private IP addresses (i.e., NAT is transparent to the client). NAT is almost independent of the application protocol transported by the IP packets. Just the protocols that transport the source address in the data field of the packet present some problems for NAT operation [9]. When using private IP addresses with NAT one must consider that all mobile hosts will appear to odier Internet hosts as being the "same computer", because they will use the same IP address (the public IP address of the NAT interface). Therefore, hosts with private IP addresses can't be used as servers, because there's no way to initiate a connection with them. This limitation can be somewhat minimized by implementing reverse NAT for selected customers. A reverse NAT is implemented by mapping specific TCP or UDP ports to specific private IP addresses. This approach, however, is not completely transparent because the external client must know the TPC/UDP port used for mapping in advance. Therefore, in practice, any application that requires opening a connection with a mobile host won't be supported in the private IP addressing configuration.

60 4

IPv6 Fundamentals

As stated in previous sections, IPv4 private addresses allow developing mobile IP networks of virtually unlimited size. However, mobile IP hosts with private addresses can act only as clients, i.e., they can't receive a message from a host they haven't initiated a connection before. To overcome this limitation, this section will evaluate the use of Internet Protocol Version 6 (IPv6) [2] addresses as an alternative to the private IPv4 addresses approach. IPv6 is considered a real alternative for implementing mobile IP networks because it can support virtually unlimited number of devices by introducing a 16-byte address space. However, IPv4 and IPv6 are not compatible. That means that adopting IPv6 will require changes in the existing infrastructure, including end user's and network devices. To allow an immediate use of IPv6 over existing IPv4 infrastructure, tunneling techniques can be employed [12,14]. Even though the IPv6 address space is not fully defined, IANA 1 has already reserved a block of IPv6 addresses for dynamically tunneling IPv6 packets over an IPv4 network [10,11]. This block is referred as "6to4 addressing scheme". Figure 2 shows the 6to4 scheme address format. SLA ID: Site Level Aggregation Identifier Interface ID: Link Level Host Identifier V4ADDR: IPv4 Address of 6tc4 Tunnel Endpoint

2002

V4ADDR

SLA ID

INTERFACE

Figure 2. The 6to4 scheme address format.

The 6to4 scheme permits the creation of dynamic tunnels intended to transport IPv6 packets over an existing IPv4 infrastructure. To implement the 6to4 scheme, an IPv6 network must have at least one public IPv4 address, referred as V4ADDR. The V4ADDR is the IPv4 address of the router interface that connects the IPv6 network to the Internet. The other router interface, attached to the IPv6 network, has an IPv6 address. This router (referred as IPv6to4 router) must support the 6to4 addressing scheme by dynamically tunneling the IPv6 packets forwarded to the Internet [14]. The tunneling technique consists in encapsulating an IPv6 packet in an IPv4 packet using the IPv4 protocol type 41, as defined in the Transition Mechanisms, RFC 1933 [13]. The IPv6to4 router can discover the 6to4 tunnel endpoint by checking the V4ADDR field from the IPv6 destination address. This enables the router to dynamically create the tunnel without previous configuration. The 6to4 tunnels are stateless, because all information required for creating the tunnels are extracted from the packets.

1

The Internet Assigned Number Authority

61

5

TRANSPARENT IPv6

This section presents a framework for assigning IPv6 addresses to IPv4 devices without any hardware or software modification. This approach is called in this paper "Transparent IPv6" (TIP6, for short). The main idea is to provide all the benefits of IPv6 addressing without introducing changes in the existing IPv4 infrastructure. In this approach, a host with a private IPv4 address will be mapped to an IPv6 address, according to Figure 3. The IPv6 address belongs to the block 2002::/16, in accordance with the 6to4 scheme. The interface identifier field is defined using the IPv4 address as the lower order 32 bits. Even though there is no restriction imposed to the identifier in the TIP6 approach, the higher order 32 bits of the interface identifier are kept zero to conform to existing standards [13]. For most implementations, SLA ID can also be kept equal to zero. The SLA ID field will be required only for very large networks that have overlapped private IPv4 address space. 2002:V4DDR::IPv4Address =

1

16

32

16

32

2002

1 V4ADDR 1

0000

OOOOOOOO

32 I IPv4 Address 1 J

SLA ID

Interface Identifier

Figure 3. IPv6 address mapping to IPv4 addresses in the transparent IPv6 (TTP6) approach.

The Transparent IPv6 approach is implemented by combining the tunneling techniques described in section 4 with one to one mapping between IPv4 and IPv6 addresses. The full approach is illustrated in Figure 4. The figure shows two IPv4 and one Ipv6 network interconnected through the Internet. The wireless representation in the local networks has been omitted. Observe that the local networks are implemented using IPv4 private addresses. IPv4 Foreign Address Pool

IPv6to4 dynamic mapping table

17Z1H.U.1

172.16.0.1 = 2002:C811:6201::OAOO:0002 172.16.0.2 = 2002:0901:0203:: ACOR0002

2002:C901:0203::AC0F:0002 Host* cde.com

192.168.0.2 (CO.A8.00.02) Host1.abc.com

192.1680.1 (C0.A8.00.01)

V4ADDR C8.C0.78.01

10.0.0.1 (0A.00.0O.01)

Figure 4. Transparent IPv6 (TIP6) Framework (Example).

The key element in the TIP6 framework is the "Transparent IPv6to4 Gateway" (TIPG, for short). The TIPG is a dual home host developed specifically for supporting our proposal. One interface of TIPG is configured with a private IPv4 address. The other interface must have a public registered IPv4 address, used as the

62

V4ADDR for creating dynamic tunnels. In the TIP6 approach, IPv4 hosts are addressed using Fully Qualified Domain Names (FQDN), i.e., IPv4 hosts must have their names registered in a DNS server using IPv6 " AAAA" records. The host names are mapped to IPv6 address using the private IPv4 address as interfaces identifier, as shown in Figure 3. TIPG permits that an IPv4 host communicates with an IPv6 host without any hardware or software modification. IPv4 hosts must be configured to use the local TIPG as both, the default gateway and the DNS server. The TIPG is not a simple NAT server. It is responsible to dynamically assign temporary IPv4 addresses to represent IPv6 hosts in foreign networks. TIPG works as a reverse NAT by mapping the destination IPv6 address to a dynamically assigned private IPv4 address that temporarily represents an IPv6 host in a foreign network. The IPv6 host can be a "true IPv6 host" or an IPv4 host with an IPv6 address mapped through the TIP6 approach (see Figure 4). To configure the TIPG, the network administrator must define an "IPv4 Foreign Address Pool", with private IPv4 addresses that are not used in the internal network. Because the mapping is not permanent, the pool size defines the number of concurrent sessions that can be handled by the TIPG. IPv6to4 dynamic mapping table

IPv6to4 dynamic mapping table

IFV4Js = R/6_s

IFV4Jc = IFV6_c

Client Request IFV4_kJ 1 IR/4_fs 1 paytoad

1c

-K

|V4AfjDR_c|V4ABbR_s|IPvri_c|IPvri_s| paybad 1 •—•

£1

IR/4_ic

lfV4_ia • IWS - j

1

1

f

1

4 RouterW

f

jT

>L

AH

r***R

V4ADDR c

l

R o ^ J—>

• IFV4_fs 1 IFV4_ic 1 paytoad

F V 4 J c | IFV4_is 1 paybad 1

• V4ADDR s

Server Response

^ - , |V4ADbR_ S |V4AbDR_c|lPv^_ s |lPv6_c| paybad 1 A - i

1 C

1

^i

FW_is 1 lrV4_fc 1 paytoad 1

Figure 5. NAT and tunneling in TIP6.

When TIPG receives a DNS query from an internal client, it maps a temporary IPv4 address to represent the foreign IPv6 server. This IPv4 address is returned to the internal client that builds a simple IPv4 packet and delivers it to the default router (TIPG). TIPG then perform two sequential operations upon the packet. First it converts the packet into an IPv6 packet using one-to-one NAT approach. Second, it tunnels the packet adding an IPv4 header. These operations are illustrated in Figure 5. The symbols used in Figure 5 are explained in Table 1.

63 Table 1. Legends for Figure 5. IPv4 ic: IPv4 fs: IPv6 s: IPv6_c: IPv4 is: IPv4 fc: V4ADDR_c: V4ADDR_s:

Internal client private IPv4 address. IPv4 address temporarily mapped to the IPv6 address of the foreign server. IPv6 address of the foreign server (found by DNS resolution). IPv6 address of the internal client (built by combining the IPv4 address and V4ADDR). Internal server private IPv4 address. IPv4 address temporarily mapped to the IPv6 address of the foreign client. IPv4 address of the TTPG in the client's network. IPv4 address of the TTPG in the server's network.

At the foreign network a similar operation happens. The TIPG in the destination network de-tunnels the incoming packet and performs a NAT operation, building an IPv4 packet with an IPv4 private address temporarily mapped to the client's IPv6 address. The reply from the server is also illustrated in Figure 5.

1. Initiate a communication with a foreign server. 2. Answer a request from a foreign client or deliver a packet to a foreign server when the FQDN is already resolved.

Find IPv6 address in the mapping table and build the IPv6 packet. Tunnel the packet with the V4ADDR header. Forward the packet to the IPv4 router.

Figure 6. Procedure for transmitting a packet in TIP6 .

Figure 6 describes the procedure in the TIP6 approach when an IPv4 host sends a packet to a foreign IPv6 host. There are two situations to consider: the internal IPv4 is a client initiating a communication with an IPv6 foreign server; or the internal IPv4 host is a server answering a request from an IPv6 foreign client. Observe that in this last situation, the mapping table is updated when the requisition arrives from the foreign host. Figure 7 describes the procedure when an internal

64

IPv4 host receives a packet from a foreign IPv6 host. Again, there are two situations to consider: the foreign IPv6 host is a client initiating a communication with an IPv4 local server; or the foreign IPv6 host is a server answering a request to an internal IPv4 client. Observe that in this last situation, the mapping table was updated when the internal client resolved the FQDN of the external server, as described in the procedure for transmitting a packet in Figure 6.

Select an IPv4 address from the Foreign Address Pool. Map the IPv4 address to the source IPv6 address in the dynamic allocation table.

Build the IPv4 packet using the IPv4 address mapped to the IPv6 source address.

I

Forward the packet to the IPv4 host.

Figure 7. Procedure for receiving a packet in T1P6 .

IPv4 packet

Receive the IPv4 packet.

c

TTP6 can be used to deploy very large wireless networks by interconnecting WLANs without any IPv4 provisioning. As seen in the previous sections, the TIP6 approach permits to assign IPv6 addresses to IPv4 hosts without modifying existing user's devices or network elements. TIP6 is implemented in a wireless network by connecting a "Transparent IPv6to4 Gateway" (TIPG) to the home agent, as show in Figure 8. In this configuration the TIPG must be configured as the DNS server for the mobile devices, but not the default gateway. The default gateway is the foreign agent, as specified by the Mobile IP standard. The TIP6 approach requires an additional route to be configured in the home agent. This route directs all packets that carry an IPv4 belonging to the foreign address pool to the TIPG. All other

65

packets can be delivered directly to the firewall. Observe that the standard NAT approach can still be used to permit the mobile devices to access IPv4 networks that do not implement the TIP6 approach. FA: Foreign Agent HA: Home Agent FN: Firewall and NAT IPv4 Router I TIPG

TIPG

~T" Figure 8. TIP6 in a mobile IP network.

Observe that TIP6 permit that independent WLAN networks use overlapped private IPv4 addresses. Therefore, it is not required any type of previous agreement between the WLAN administrators in order to support intercommunication. 6

Conclusion

This paper presented a novel approach to manage the problem of provisioning IPv4 addresses for Mobile IP implementation. This approach is called "Transparent IPv6" (TIP6, for short). The main idea is to provide all the benefits of IPv6 addressing without introducing changes in the existing Mobile IPv4 infrastructure. This approach could be immediately employed to deploy very large wireless networks by interconnecting WLANs without any IPv4 provisioning. We have shown that the TIP6 approach extends the standard NAT functionalities by permitting a mobile device to receive connections from other mobile or fixed hosts connected to the Internet. An IPv4 host in a TIP6 network can communicate with other TIP6 networks or with "true" IPv6 hosts. This feature will permit the coexistence of emerging IPv6 Mobile IP implementations with existent IPv4 Networks. TIP6 is still in its early stages of development. Future works will present results concerning scalability and performance (latency introduced by the TIP6 approach). Also, we intend to explore QoS and security features offered by IPv6 to create virtual private communities of mobile IP subscribers.

66

References 1. C. Perkins, ed., IP Mobility Support, IETF RFC 2002, Oct. 1996. 2. S. Deering and R. Hinden, Internet Protocol, Version 6 (IPv6) Specification, IETF RFC 1883, Dec. 1995. 3. J.B. Postel, ed., Internet Protocol, IETF RFC 791, Sept. 1981. 4. D. Johnson and C. Perkins, Mobility Support in IPv6, ACM Mobicom 96, ACM, Nov. 1996, pp. 2737. 5. GSM Association. An Overview of GPRS (General Packet Radio Service). http://www.gsmworld.com/technology/gprs.html 6. Y. Rekhter, B. Moskowitz, D. Karrenberg, G. J. de Groot , E. Lear, Address Allocation for Private Internets, IETF RFC 1918. 7. Kim Fullbrook, GSM Association, Tackling the Mobile Addressing Problem. GPRS sessions at the 3GSM World Congress 2001. http://www.gsmworld.com/technology/gprs_presentations.html 8. C. Perkins, ed., IP Mobility Support for IPv4, revised, draft-ietf-mobileiprfc2002-bis-03.txt. 9. Michael Hasenstein, IP Network Address Translation. http://www.suse.de/~mha/HyperNews/get/linux-ip-nat.html 10. R. Hidden, S. Deering. IP Version 6 Addressing Architecture. IETF RFC 2373, July 1998. 11. R. Hidden, M. O'Dell, S. Deering. An IPv6 Aggregatable Global Unicast Address Format. IETF RFC 2374, July 1998. 12. B. Carpenter, C. Jung. Transmission of IPv6 over IPv4 Domains without Explicit Tunnels. IETF RFC 2529, March 1999. 13. R. Gilligan, E. Nordmark. Transition Mechanisms for IPv6 Hosts and Routers. IETF RFC 1933, April 1996. 14. B. E. Carpenter, K. Moore, B. Fink. Connecting IPv6 Routing Domains Over the IPv4 Internet. Cisco Internet Protocol Journal, Volume 3, Number 1, March 2000.

Part 3 Error Control and Mobile Applications

A LOW COST ERROR CONTROL SCHEME FOR HARD REAL-TIME MESSAGES ON WIRELESS LANS JUNGHOON LEE Dept. Computer Science and Statistics E-mail: jhlee @ venusl. cheju. ac. kr YOONJOON LEE, SIN KIM Dept. Nuclear and Energy Engineering E-mail: {leeyj, [email protected] This paper proposes and analyzes a low-cost error control scheme for hard real-time messages in wireless local area networks. As a transmission control based real-time communication, the bandwidth allocation scheme decides the size of repetition cycle and the capacity vector for the given stream set, so that each message meets its time constraint. The proposed error control scheme makes the receiver node report errors in a best-effort manner during the contention period. It also makes the sender retransmit the damaged message via the overallocated bandwidth on the contention free period. This scheme obviates the interference of error control messages to the other real-time messages. The performance measurement result executed via simulation using SMPL shows that, with the proposed error control scheme, the hard real-time network can be built at cost lower than, but performance comparable to the dual link network.

1

Introduction

Wireless local area networks(WLANs) are emerging as an attractive alternative, or complementary, to wired LANs, because they enable us to set up and reconfigure LANs without incurring the cost of wiring[l]. WLANs are expected to be the solution to the problem of meeting the growing demand for mobile clients to have access to the existing high-speed wired networks. As the need for broadband multimedia communications involving digital audio and video grows, it is increasingly important for communication systems to support the time-constrained traffic. Furthermore, in the hard real-time system built upon the WLAN, such as factory process control system, the exchanged messages are typically critical or essential for the proper operation of the entire system. We mean by real-time that an information package has to be received by the recipient before a certain deadline [2]. The traffic requires bounded delays, but is usually tolerant of some message loss. Hence, the requirement of these messages includes a guarantee from the system that they should be delivered within their deadUnes as long as there is no network error[3]. However, wireless transmissions are subject to interference from outside sources, absorption, scattering, and fading. While the error correction code such as

69

70

FEC(Forward Error Correction) is able to recover many damaged messages so reduce the probability of message discard, it can not cope with all of the transmission errors. Though this error correction mechanism seems to be promising, information overhead and processing complexity are not negligible. More errors can be recovered with the retransmission of damaged messages. For a real-time system, the traditional error recovery schemes, proposed for non-real-time data transmission, may introduce an unacceptably long delay. In addition, the retransmitted message may interfere or prolong the delivery of other normal messages, resulting in the cascaded deadline miss. Hence, the error control scheme on real-time system should consider the time constraint of messages and recover as many errors as possible within the deadline of message. The networking community has explored a broad spectrum of solutions to deal with wireless error environments. They range from local solutions that decrease the link error rate observed by upper layer protocols to transport protocol modifications that modify the behavior of the high level protocols[4]. The local error control scheme fits for real-time communication as the end-to-end retransmission demands bandwidth on every link, while the local retransmission requires extra bandwidth only where it is truly needed. In this paper, we propose the local error control scheme for real-time messages on the wireless LAN. The proposed scheme is built on top of a transmission controlbased message scheduling policy. It makes the receiver node report errors in a besteffort manner via the network time reserved for the non-real-time traffic. It also makes the sender retransmit the damaged message via the overallocated bandwidth. The proposed scheme completely obviates the interference of error control messages to the other real-time messages. This paper is organized as follows: After issuing the problem in Section 1, we will introduce the related works in Section 2. Section 3 and 4 describe the bandwidth allocation scheme for the real-time message stream and the proposed error control scheme in detail, respectively. Section 5 shows the performance measurement results and Section 6 summarizes and concludes this paper. 2

Background and Related Works

The IEEE 802.11 standard is commonly accepted for wireless local area networks[5]. It defines the physical layer and the medium access control in the data link layer for wireless local area networks. For both real-time messages and nonreal-time messages, the IEEE 802.11 standard provides the means to realize contention free and contention based medium access over a single physical channel. In particular, it specifies two alternating phases of medium access control: contention period with DCF(Distributed Coordinated Function) and contention free

71

period with PCF(Point Coordinated Function) as shown in Figure 1. The PCF relies on the PC(Point Coordinator) to perform polling, enabling polled stations to transmit without contending for the channel. The method by which polling tables are maintained and the polling sequence is determined, is left to the implementor. CFP repetition CF B

PC

C

'

DCF

Figure 1. Time axis of IEEE802.11 protocol

M. Mock and E. Nett have proposed polling, or access, mechanism based on TDMA(Time Division Multiple Access) for contention free period[l]. The scheduling mechanism assigns fixed size slots to stations, a station holding a time slot has the exclusive right to send messages during that time slot. Time slots allocated for sending messages must comprise the worst case time needed to send the longest possible message. TMDA is basically designed for periodic message stream transmission in static environments, thus a suitable approach for guaranteeing hard real-time communication. While collisions are avoided and the communication delay becomes predictable, the pure TDMA approach results in a poor utilization of network bandwidth due to the fact that the slot size is fixed. As an approach for the dynamic environments, a unified wireless architecture has been proposed[6]. The authors addressed how to support real-time and non-realtime communication services in a wireless LAN with D-TDD(Dynamic Time Division Duplex) transmission. For real-time traffic, a non-preemptive EDF(Earliest Deadline First) policy is adopted. The polling order is determined based on the contracted traffic characteristics so that every node is polled at the regular interval once every period of message stream. However, the size of its polling sequence table may increase, let alone the admission control mechanism becomes too complex. As for the error control scheme for real-time traffic on wireless LANs, H. Bengtsson and his colleagues have proposed a protocol based on retransmission fulfilled on demand within a given time window[7]. Each retransmission is coded with a varying number of redundant symbols. The protocol is supported by the Reed-Solomon codes that can transform the message into a data stream with the desired probability of success. There remains some margin to improve the recoverability if some supplementary error control mechanism is integrated at LLC(Logical Link Level) layer. In addition, a two-step adaptive error recovery scheme is proposed which uses both Reed-Solomon code and Stop-and-Wait ARQ(Automatic Retransmission reQuest)[8]. The protocol is based on both the wireless channel condition and the deadline constraint, adaptively selects the best error correcting code by looking up

72

an optimal code table which is predetermined before the start of message stream. As every transmitted packet needs one or more ARQ message according to Stop-andWait scheme, the error control induces unnegligible overhead, so reduces the guarantee ratio. 3

3.1

Bandwidth Allocation

Network and message model

BSS(Basic Service Set) is the fundamental building block of the wireless LAN architecture. It is defined as a group of nodes that are under the direct control of a single coordination function. All nodes in a BSS can communicate directly with all other nodes in the same BSS. In a BSS shown in Figure 2, AP(Access Point) supports range extension by providing network connectivity between BSSs and coordinates the message transmission within a BSS[5].

Figure 2. Network architecture.

According to the general real-time message model, we assume that there are n streams of real-time messages, 57, S2 , ... , Sn. Each stream can be characterized as follows: A message arrives at the beginning of its period and must be transmitted by the end of the period[9]. The period of a stream Si is denoted as Pi and the maximum length of Si is Ci. The first message of each stream arrives at time 0, namely, the start of the first repetition cycle. 3.2

Motivation

As real-time messages are deterministic, that is, all of their timing characteristics are known in priori, most of the research in this area has followed the guarantee strategy. The goal has been to guarantee that all real-time messages will meet their deadlines during error-free operation of the network. Guaranteeing schemes for multiple-access network such as wireless LAN can be classified into two categories, namely, access arbitration and transmission control approaches. The first concentrates on determining when a node can send a message over the shared channel, for example, D-TDD described in Section 2. The admission control function on this network is based on the rate monotonic algorithm, however, it can be only approximated when using priority driven protocols. As N. Malcomn has

73

pointed out, access arbitration schemes suffer from unnegligible priority arbitration time as well as dependency on the transmitted packet size[9]. On the contrary, the second focuses on deciding how long a node can continue to send messages over the channel. This scheme is basically designed for the transmission of periodic message stream in static environments. The scheme may bring bandwidth waste during the guaranteeing procedure. However, if there is not a considerable difference between both period and transmission time of the message streams, the scheme may reduce the bandwidth waste. Moreover, the wasted bandwidth could be used for the purpose of error control, which is the essential part of the communication over the wireless networks. 3.3

Allocation Scheme

With a transmission control based real-time communication in the wireless LAN, AP polls each node once a contention free period in the round-robin fashion. Bandwidth allocation procedure determines the size of contention free interval as well as the time a node can send its message when polled. Namely, if we denote the bandwidth allocation scheme as B, the following relation holds. B : (Pi, Ci) • F, Hi (1) , where F is the size of one repetition cycle and Hi is the maximum time during which node Si send its message when polled. Repetition interval(F) Hi

H2

H3

r

Hn

C.T

Polling overhead

Figure 3. Bandwidth allocation As shown in Figure 3, each node is polled once a repetition interval whether it has messages to send or not, and is able to transmit for at most Hi. There exist a number of bandwidth allocation schemes which correspond to the transmission control based real-time communication^]. Among these, we choose local allocation scheme fox simplicity, as we are mainly concerned on the error control scheme. It is natural that another allocation scheme may be applied. If we let U denote me utilization of message stream of given stream set, it can be calculates as 2 (Ci /Pi). In addition, y denotes the polling overhead in a repetition cycle. Then die local allocation scheme calculates the network parameters as in Eq. (2).

74

- 3 + y/9 + SP^ 2

ly

H, = - ^ - 5 . (jTRT - y) . for all i

(2)

, where Pmin is the smallest period in the stream set. If that allocation meets the constraints produced from the given stream set, it is called as a feasible solution. Then the duration of contention period, CP, is the remaining portion of F, as in Eq.(3). CP = F-XHi (3) 4

4.1

Error control scheme

Error report

If the error correction code fails to recover the error, additional explicit messages should be sent for retransmission request or error report, which demands additional bandwidth allocation[10]. As other error control makes the receiver respond to every received packet, it wastes the network time[7,8]. This waste can be reduced by making the receiver send the error report only when the sender completes transmission of entire packets. It is desirable to send this message via the contention period in order not to interfere the transmission of normal real-time messages. To construct the error report message, the receiver initializes the error packet list when the first packet arrives. If the sequence number of this packet is not 1 but k, the receiver adds the numbers from 1 up to (k-1) into the list. From then, the receiver appends each number of the erroneously received packet. As the receiver is also polled and knows the total number of packets transmitted in a period, it can decide when to send error report message to the transmitter. Set counter with 3

[fl I p i f21 p

f3

P

—

Error report fu

P

CP

• Time Axis Figure 4. An example of error repotting

Figure 4 shows the example, where a message is transmitted with u packets and the receiver knows it from the connection establishment. At the initial state, a packet numbered as 3 (not 1) arrives. Then die receiver initializes and adds 1 and 2 to error list as well as sets counter to 3. The counter increases each time the receiver is

75

polled. When the counter reaches to u, receiver sends error report back to the sender via contention period, if possible. After all, error list is reported to the sender periodically as long as there is additional available network time as in SNR error control scheme [10]. 4.2

Error Control

Deferred or erroneously-transmitted real-time packets must be scheduled again before their deadlines only when the sender is polled redundantly to eliminate the interference to other message transmission. For the given stream set, the parameter selection procedure, called as bandwidth allocation, decides the F and capacity vector after the analysis of the stream set. The procedure first searches the period that has the minimum available transmission time, that is, which receives the smallest number of poll message. To meet always the time constraint for a stream means that the message transmission on such a period meets time constraint. The periods, however, other than that has the minimum transmission time, have extra bandwidth. Retransmission via this overallocated bandwidth does not result in missing deadline nor interfere the transmission of normal messages. paCket poU

'

^

Expire/reXmit req

Req reXmit

done/Ack

oll/reXmit

Expire/reXmit req (a) receiver side

Poll/Xmit

(b) sender side Figure 5. State diagrams for error control

Figure 5 shows the state diagrams of sender and receiver, respectively. As shown in Figure 5(a), when the receiver receives the first packet in the initial state(State 1), state is changed into State 2, setting the counter with the number of packet as well as initializing the error list. State 2 continuously receives poll message or packet, updating the counter and error list when needed. When the counter expires and all of packets are correctly received(done), state is changed to State 1. Otherwise, retransmission request message is generated and the receiver goes to State 3. The receiver tries to send the message via the contention period. In State 3, the receiver waits for the retransmitted message until all of packet errors are recovered or the first packet of a new message arrives. Figure 5(b) shows the state diagram of sender side. A new message arrival makes state transition from State 1 to State 2. In State 2, the sender transmits the packets one by one each time it is polled. If an acknowledgement message arrives, the sender goes to State 1 and wait for the new

76

message. Arrival of a retransmission request message changes the sender to State 3, where the sender transmits again the requested packet via overallocated bandwidth. If a sender is not additionally polled in a period, it abandons the retransmission.

5

Performance Evaluation

We measure the performance of the proposed error control scheme via simulation using SMPL[11], which is a general purpose portable discrete event simulation library written in C. It uses an event scheduling oriented view so provides a convenient and powerful protocol simulation. This section exhibits two results, success ratios according to bit error rate and utilization of stream sets. The success ratio is calculated by dividing the number of messages transmitted successfully within their deadlines by the total number of generated messages. As a performance reference, we choose the dual link network model. In this model, all the duplicates are transmitted, irrespective of whether or not the earlier duplicate suffered delivery errors or not[12]. The receiver cannot recover the error only if all of two duplicates experience the transmission error simultaneously. Though it needs twice as much bandwidth as the single LAN, it is the upper bound for the performance of error control scheme.

Figure 6. Success ratio vs. error rate

Figure 7. Success ratio vs. utilization

For the first experiment, we have generated 40 stream sets that have arbitrary number of streams individually. The utilization of each set ranges from 70 % to 79 %. This experiment measures the success ratio for the bit error rate from 10"4 to 10"8. As expected, the message transmission supported by the proposed error control scheme outperforms that without error control scheme for all error rates as shown in the Figure 6. The graph shows that the proposed scheme enhances success ratio up to as much as 50 % at maximum when error rate is 10"5. In addition, it is natural that

77

the success ratio will increase as the utilization of stream set goes lower because each stream has more bandwidth available for error control. Therefore, the second experiment has measured success ratio changing utilization of stream sets from 0.5 to 0.8 with error rate fixed to 10~5. The stream set above utilization 0.8 is not considered because the bandwidth allocation scheme cannot find feasible capacity vector. As shown in Figure 7, the success ratio enhancement decreases as utilization increases as expected. In addition, analysis result approximates the error rate, Ei, for a stream Si. To begin with, let u denote the number of packets for a message and v denote the number of recoverable packets, respectively. Then u can be calculated as [d/Hi], while v as {.Ri/Hij, where Ri is the average redundant bandwidth for each Pi. Ri is calculated as Eq. (4). Ri = (l-U)Pi-j (4) With the previously defined symbols Ei can be approximated as Eq. (5). Ei=l--LuCkeika-eifka-edk

(5)

, where e, is the packet error rate. The detailed description as well as the comparison with the simulation results will be published at another paper. 6

Conclusion

This paper has proposed and analyzed an error control scheme for wireless LAN networks according to the transmission control based scheduling policy in hard-real time communication. The bandwidth allocation scheme decides the size of repetition period, size of contention period, and capacity vector. The coordinator polls each node one by one following the round robin fashion, while the node, when polled, sends its message up to the amount allocated to it during the allocation procedure. The proposed error control scheme makes the receiver node report errors in a besteffort manner through the contention period that is originally reserved for non-realtime traffic. It also makes the sender retransmit the damaged message when the sender has additional bandwidth in a period. This scheme obviates the interference of error control messages to the other undeferred real-time messages. Using the proposed error control scheme, the hard real-time network can be built at cost lower than, but performance comparable to the dual link network. References 1. M. Mock, E. Nett, "Integrating hard and soft real-time communication in autonomous robot system," IEICE Trans. Communication, pp. 1067-1074, May 2000.

78 2. James F. Kurose et. al., "Real-time communication in packet-switched networks," Proceedings of IEEE, Vol. 82, No.l, pp.122-139, Jan 1994. 3. K. Arvind, Krithi Ramamritham and John A. Stankovic, "A local area network architecture for communication in distributed real-time systems," Journal of Real-Time Systems, Vol. 3, pp.115-147, May 1991. 4. D. Eckhardt, P. Steenkiste, "Improving wireless LAN performance via adaptive local error control," Proc. IEEE ICNP, 1998. 5. B. Crow, I. Widjaja, J. Kim, P. Sakai, "IEEE 802.11 wireless local area networks," IEEE Communication Magazine, pp.116-126, Sep 1997. 6. S. Choi, K. Shin, "A unified wireless LAN architecture for real-time and nonreal- time communication services," IEEE/ACM Trans, on Networking, Feb. 2000. 7. H. Bengtsson, E. Ulemann, P. Wiberg, "Protocol for wireless real-time systems," Proc. Euromicro Conference on Real-time Systems, 1999. 8. D. Qiao, K. Shin, "A two-step adaptive error recovery scheme for video transmission over wireless networks," IEEE INFOCOM'00, Tel Aviv, Israel, pp.1698-1704, Mar 2000. 9. Nicholas Malcolm and Wei Zhao, "Hard real-time communication in multipleaccess networks," Journal of Real-Time Systems, pp.37-77, 1995. 10. A. N. Natravali, W. D. Roome, K. Sabnani, "Design and implementation of a high-speed transport protocol," IEEE Trans. Communication, Vol. 38, No. 11, pp.2010-2024, Nov 1990. 11. M. H. MacDougall, Simulating Computer Systems: Techniques and Tools, MIT Press, 1987. 12. H. Kopetz, "TTP-A protocol for fault-tolerant real-time systems," IEEE Computer, pp.14-23, Jan 1994.. Acknowledgement This work has been carried out under the Nuclear R&D Program by Ministry of Science and Technology (MOST), Korea.

B U N D L E S R E P L A C E M E N T IN GATEWAYS

Electrical

and Computer

BEIZHONG CHEN Engineering Department, Rutgers NJ 08854, USA E-mail: [email protected]

University,

Piscataway

Computer

Science

KHALED ELBASSIONI Department, Rutgers University, Piscataway E-mail: [email protected]

NJ 08854,

USA

Panasonic

IBRAHIM KAMEL Information and Technologies Laboratory, Princeton, E-mail: [email protected]

NJ 08540,

USA

This paper studies how to manage limited memory available in home gateways like STB. In the OSGi model, services are implemented in software bundles that can be downloaded separately from the Internet and executed in the gateway. OSGi specifications define a service dependency scheme that maps the relationship between services. The problem we solve is: when the memory is full, which service(s) will be stopped or kicked out of memory to start a new service. Note that stopping a given service means that we stop all the services that depend on it. Our goal is to minimize the total number of stopped services. Because of the service dependencies, traditional memory management techniques, such as LRU, best fit, first fit, worst fit are not suitable. This paper presents an optimal algorithm for the above service replacement problem. We also present a more efficient heuristic in terms of memory and running time. Preliminary experimental results are presented to evaluate the performance of the proposed algorithms.

1

Introduction

With the widespread of the broadband to homes, the Internet is becoming more and more popular inside the homes. The Internet is used not only for connecting computing devices such as PCs and PDAs but also for home appliances like TV, DVD, washers etc. Merloni Elettrodomestici, an Italy based company announced its Internet washer Margherita2000 7 that can be connected to the Internet through a modem. The manufacturer or the owner can program the Internet washer remotely. Matsushita Electric showed its Panasonic Internet enabled microwave during the 2001 Consumer Electronic show. Cooking recipes and heating instructions can be downloaded to the microwave. It is expected in the near future that all the appliances in the house will be connected to the Internet. Remote diagnosis and configuration of the appliances is one of the many advantages that the consumer can gain from Internet 79

80

appliances. In the entertainment domain there are several interesting applications. Users can download movies on demand and an Electronic Programming Guide (EPG) that contains the TV schedule. Power companies are keeping an eye on home networking because it will allow them to provide value-added services such as energy management, remote measurement, and better power balance that reduces the likelihood of blackout. There are several initiatives to define the specification for API and the protocol for networks inside the home, like HAVi 2 , U P n P 5 ' 6 , Jini 4 , HomeRF 3 , to name a few. It is expected that multiple home networking will coexist in the home and inter-operate through the home gateway. The home gateway is in action also as a single point of connection between the Internet and the home. OSGi 8 ' 9 (Open service Gateway initiative) is a consortium of companies committed to define common specifications for the home gateway. According to OSGi model, the gateway can host services to control and operate home appliances. There is not consensus among the industry on whether the gateway will be a separate box or it will be integrated in home appliances like DTV or STB (Set-Top-Box), or whether the gateway functionality will be centralized in one device or distributed among several appliances. However, the gateway will generally have limited resources, especially main memory. Main memory is required to run system administration, services and applications for home devices. This problem will be more severe if the gateway functionality is implemented in home devices like STB or TV. This paper discusses memory management in gateways and proposes new algorithms for managing services and bundles in a limited memory environment. We address this problem in the context of the software architecture proposed by OSGi. Memory management in computers has been an active area of research for several decades 10 . This research spans multiple disciplines such as operating systems, database management to name a few. In the OSGi model, services are implemented in software bundles that can be downloaded separately from the Internet and executed in the gateway. For example, HTTP service is implemented in a bundle. Other applications like home security, energy management, and Internet game each can be implemented in a separate bundle. Bundles can depend on each other; for example, a home security bundle uses an HTTP service to provide external connectivity. Bundle and service management in gateways differs from traditional memory management in the following aspects: • The memory unit in traditional memory management techniques is the disk page, whereas the memory unit in a gateway application is the memory required to execute the bundle.

81

• Traditional memory management techniques assume that disk pages are independent", while bundles may depend on each other as explained in Section 2. • Bundles might have priority levels associated with them that are determined by the importance of the corresponding applications. • Kicking a bundle from the memory results in aborting the service, while kicking a page from the memory costs one I/O. 2

Bundle model and bundle dependency in OSGi

Figure 1 shows the software architecture of a gateway according to the OSGi model. Each service or set of services is implemented in a software container called a bundle. The framework handles the basic bundle management functionality. Our proposed algorithm is implemented as a part of the framework. Applications are designed as a set of services, with each service implementing a segment of the overall functionality. These services and other extension services are then packaged into a bundle. The gateway can download the corresponding bundle when it becomes necessary. In order to share its services with other bundles, a bundle can register any number of services to the Framework. The following terminologies are used throughout the rest of the paper: • Bundle: the functional and deployment unit for shipping services, it contains implementation for any number of services. • Service: a self-contained component that performs certain functionality. • Service Instance: the execution thread of a service in the Framework. A bundle may import services provided by other bundles. We use a dependence graph to model this dependency among services. Formally, given a set of service instances S = {si,...,sn}, which are currently resident in main memory, let G(S,E), be a directed acyclic graph with vertex set S and edge set E, describing the dependence among these instances. There is a directed edge (si, Sj) € E if and only if Sj depends on Sj. Since it is natural to assume that each application instantiates its own copy of a given service, the dependence graph will consist of a forest of rooted trees, where each tree represents the service instances instantiated by a given application as shown in Figure 2. "although, depending on the application, disk pages might be dependent on each other, for examples, two pages belong to one loop

82

O

C/3

HTT

UH

an

O

a<

an c= nn
c*i C/3

o

V2


tL>

s

bo

fied

vie



Application Bundles

-J

H

o

,"£i B O

^ >> curit

Basic Bundles

0}

Framework Java Virtual Machine (pJava,J2ME,etc) Operating System/RTOS Hardware Figure 1: The software architecture of a gateway in the OSGi model. 3

Problem definition

The gateway might need to free memory space to accommodate a new service that is triggered by connecting a new device to the network or upon explicit local or remote request. Although the amount of memory required to execute a service is changing with time, the application service provider (or the author who provides the bundle) can give approximate estimates such as average and maximum amount of memory required to execute the services in this bundle. Memory space might be requested by a running service instance that requires more memory to complete its service. At this time the gateway framework has to pick a victim service instance or instances to terminate to fulfill the request. Note that the victim service instance might be the requesting service itself. The proposed algorithm has to fulfill the following criteria: • The number of victim service instances should be minimized. • The victim service instance is the least important among the running ones. • The algorithm should be fast enough and does not require much memory. • The total amount of reclaimed memory is enough to fulfill the requested memory.

83

Qu^n-demand(30O .Internet Game(65) C2SoPlayer(50) <^Sservice(45) C^PService(25) a: Audi-on-demand application.

b: Internet game.

Figure 2: Dependence graph for two currently running applications (numbers indicate memory requirement).

This can be best described by the following simple example. Suppose that there are two applications that are running on the gateway. The first application uses audio-on-demand service which depends on the audio player service which, in turn, depends on the UDP service. The service dependency graph for the audio-on-demand is shown in Figure 2:a. The second application (Figure 2:b) is the Internet game, which has a game service that depends on the HTTP service. Now we would like to start the home security service, which requires 100 memory units. Apparently, home security service has a higher priority than both Audio-on-demand and Internet-game. Thus it is reasonable to kick at least one of the services to start the home security service. Assuming that both Audio-on-demand and Internet-game have the same importance level, it makes more sense to kick the equalizer of the audio-on-demand because it results in killing less number of service instances and fulfill the memory demand. In many applications it is possible to kick one service in the application and keep the application running. In Section 5, we propose two algorithms for service replacement for memory management in gateways. 4

Formal description of the problem

More formally, our problem can be described as follows. Let S = {si,..., sn} be the set of service instances currently resident in gateway memory. Assume that instance s; occupies M(si) storage. Let G(S,E) be the forest of trees describing the dependency among the set of instances S. For a vertex v in G, let us denote by v~ the set of vertices of the subtree of G rooted at v (including

84

v itself), and for a subset of vertices V C S, let V~ — Uv^v(v~)Given that a new service instance s, with memory requirement M(s), has to be created, it might be required to eject some currently existing instances in order to make room for the new instance. Assume that the extra required memory for this operation is Mt units, that is Mt = M(s) — Mf, where Mf is the current amount of available memory. Clearly, when a service instance is ejected, all instances depending on it will be ejected as well. Our goal is to reduce the number of removed instances. More precisely, it is desired to find a subset F C S o f minimal number of dependents, whose ejection, together with all its dependents, will make available a total memory of at least Mt units. Letting M(S') as of finding

= £ s 6 S , M(s) for any S' C S, our problem can be formulated m i n { | y - | : V C S,M(V~)

> Mt}.

(1)

There is a natural generalization of problem (1) which can be stated as follows. Suppose that different service instances differ in importance and, therefore, are assigned different priorities. In such a case, it is reasonable to assign a weight w(s) to each instance. Instances with large weights are considered more important. When it is necessary to eject some instances from memory, it desirable to reduce the number of deleted high priority instances. Thus the problem now becomes of finding min{V7(V~) | V C S,M{V~) > Mt}. This last problem, in the absence of dependencies, is closely related to the well-known Knapsack problem, which is NP-hard in general 1 . However, the Knapsack problem admits a pseudo polynomial algorithm which runs in 0(n2W) (see, for example, 1 1 ), where W is the maximum weight. We shall show in Section 5.2 that a similar result holds for our problem (1), even if in the case of dependencies among service instances. 5

Algorithms

In this section we present two algorithms for solving the gateway bundles replacement problem. The first algorithm is a simple heuristic that runs in 0(n2) time and requires linear space. The second algorithm computes an optimal solution in 0(n3) time and 0{n2) space. 5.1

Heuristic algorithm

Our heuristic greedily tries to pick, as a victim for deletion, the service instance with the smallest number of dependents, whose ejection will free the largest amount of memory. Towards this end, the heuristic will pick for deletion the service instance s which maximizes the ratio M{s~)/\s~\ of the total memory

85

to the number of dependents. This selection tends to decrease the number of deleted instances. Looking back at the example in Figure 2, we can see that the ratios M ( s ~ ) / | s ~ | for the different service instances are as follows: Audio-ondemand (52.5 = [30 + 50+ 25 +105]/4), Audio-player (37.5), UDP Service (25), Equalizer (105), Internet game (50), and HTTP Service (45). Thus the service instance with maximum ratio is the Equalizer whose removal will give enough memory to start the new service (requiring 100 memory units). Should we have used the First Fit strategy, on the other hand, we might have selected to remove the Audio-on-demand instance, which results in removing 4 instances instead of only 1. Note also that, for this particular example, the Best Fit algorithm would also remove the same instance (the Equalizer) selected by the SDP Heuristic. Step 1. For each node s £ S, compute the accumulative size and memory \s~\ and M(s~) as described by the depedence forest G. Step 2. If Mt < 0 then halt. Step 3. Pick the the node s with the maximum value of the ratio i j l i ' . In case of ties, pick the node with the smallest number of dependents \s~\. Delete s and all its depedants. Step 4. For every node u on the path from s to the root of the tree containing s, set \u~\ +- \u~\ - \s~\ and M(s~) <- M ( u _ ) - M(s~). Step 5. Set Mt*r-Mt-

M(s~) and go to Step 2.

It is easy to see that the above heuristic can be implemented in time. 5.2

0(n2)

Optimal algorithm

It is well-known that the Knapsack problem admits a pseudo polynomial algorithm. In this section, we extend this solution to problem (1). The technique used is dynamic programming. Specifically, let S = { s i , . . . ,sn} be the current set of service instances listed in breadth-first search order (that is, we first list a root, followed by its children, then by the next level of children, and so on). We shall consider incrementally the sets {s\}, {si,S2J, {«i,S2,S3}, ..., computing for each set the maximum amount of memory that can be achieved by deleting a subset of it of a given number of dependents. Note that n is a trivial upper bound on the total number of instances (or weight) that can be achieved by any solution. Since elements of S are considered in breadth first search order, when we consider an element Sj, all its ancestors must have been

86

already considered. Therefore, the amount of memory added by considering the new element s* is either 0 or M(s^). For each i £ { 1 , . . . ,n} and each w € { 1 , . . . , n}, let 5j ;lu denote a subset of { s i , . . . , s,} whose total weight is exactly w, and whose total memory is maximized. Let A(i,w) — M(S^W) if the set Si%w exists, and A(i,w) = —oo otherwise. We initialize A(i,0) = 0 for all i = 1 , . . . , n. Clearly A(l, w) is known for every w € { 1 , . . . , n}. The other values of A(i, w) can be computed incrementally using the following recurrence A(i + l,w) = max{A(i,w),M(s-+1)

+ A(i,w -

\s~+1\)},

if ki+il < w, Si+i $ S^~w, and A(i + l,w) = A(i,w) otherwise. Clearly the sets SitW can be easily obtained by a similar recurrence. The minimum weight of a subset solving our problem is then min{w | A(n, w) > Mt). Thus we get an 0(n3) algorithm for solving problem (1). The space used is 0(n2) since for each w — 1 , . . . , n we need to maintain the sets SiiW. 6

Experimental results

Preliminary experimental results comparing the performance of the two proposed algorithms are presented in this section. In our experiments, we used two different values for the gateway memory size: 64 Mbytes and 128 Mbytes. Initially, bundles are generated with random sizes and loaded into gateway memory, until the memory becomes nearly full. Then, a new bundle is created and both algorithms are run on the current set of bundles to determine which one(s) should be kicked out to make room for the new bundle. Bundle sizes are selected uniformly in the range from 100K-10000K. When a new bundle is created, it is made dependent on another randomly selected bundle with probability 0.9. To show the performance difference between the heuristic and optimal algorithms, we increase the size of the newly created bundle from 100K to 50000K. Tables 1 and 2 show our results, corresponding respectively to total memory sizes of 64M and 128M. In the tables, we show the number of kicked out bundles as the size of the newly created bundle is increased. The experiment was repeated 1000 times and the average number of kicked out bundles is measured. The last row of the table shows the performance ratio of the heuritic to the optimum. As can be seen from the tables, when the size of the new bundle is small, the heuritic performs quite close to the optimum. But as the size of the new bundle is increased, the performance gap between the two algorithms increases. This should be expected since, when the size of the newly created bundle is small, both algorithms will need to kick out only a small number of bundles. However, with larger bundle size, the number of bundles to be kicked

87 Table 1: Average number of kicked out bundles for total memory size=64M New bundle size (K) Optimal Heuristic Ratio

100 1.00 1.06 1.06

500 2.00 2.18 1.09

5000 3.71 4.88 1.32

10000 6.63 9.36 1.41

20000 11.76 17.41 1.48

30000 19.77 29.17 1.48

50000 35.77 49.47 1.38

Table 2: Average number of kicked out bundles for total memory size=128M New bundle size (K) Optimal Heuristic Ratio

100 1.00 1.02 1.02

500 2.00 2.12 1.06

5000 3.51 4.79 1.36

10000 6.26 9.13 1.46

20000 11.23 17.00 1.51

30000 18.32 28.26 1.54

50000 30.69 46.19 1.51

out might be larger, and therefore, the number of choices becomes also larger, making it easier for the heuristic to miss the optimal solution. Finally, we notice that for both gateway memory sizes, the performance of the heuristic remains within 1.6 of the optimal. Thus given the simplicity and effeciency of the heuristic, it becomes a good candidate for use in practical applications. 7

Conclusion

We have considered the problem of managing services and bundles in a limited memory gateway. We addressed the problem in the context of the software architecture proposed by OSGi. One of the main differences of our problem from traditional memory management is that the dependence among different services has to be taken into consideration. Two algorithms were proposed to solve the problem in the special case when the dependence graph is a forest of trees. Simulation results indicate that our proposed heuristic is a good candidate for use in practical environment, as its performance is close to the optimal solution. Providing a solution for the general case remains an important extension to our work. References 1. M. R. Garey and D. S. Johnson, Computers and Intractability, Freeman, New York, 1979. 2. HAVi Consortium, Havi Specification VI.0, 2000, http://www.havi.org.

88

3. HomeRF, Technical information, http://www.homerf.org/techinfo/whitepapers.html. 4. Sun Microsystems Inc., Jini Architectural Overview - January 1999, Sun White Paper, January 1999. 5. Microsoft Corporation, Universal Plug and Play Device Architecture Reference Specification, Version 1.0. 6. Microsoft Corporation, 1999, Universal Plug and Play: Background. 7. Margherita2000, The first washing machine on the Internet, http://www.margherita2000.com/sito-uk/it/home.html. 8. OSGi Consortium, OSGi Specification 1.0, 2000, http://www.osgi.org/about/specl.html. 9. OSGi Consortium, OSGi Specification Overview, 2000, http://www.osgi.org/about/Specoverview.html. 10. A. Silberschatz and J. L. Peterson, Operating System Concepts, Addison Wesley, 1989. 11. Vijay V. Vazirani, Approximation Algorithms, Manuscript, College of Computing, Georgia Institute of Technology.

Part 4 Bluetooth and 802.11

FLASH NOTES OVER BLUETOOTH WIRELESS TECHNOLOGY

PEKKA JAPPINEN, JARI PORRAS Lappeenranta University of technology, P.O.Box 20, 53851 Lappeenranta, E-mail: {Pekka. Jappinen, Jari. Porras} @lut.fi

Finland

In this paper a usage model for Bluetooth technology, i.e. Plash notes, is studied. The Flash note usage model can be defined as the problem of getting localized, upto-date information to the user's portable device according to the user's personal wishes. The usage model is further developed and a Flash GORDoN architecture containing both authentication and certificate issues are discussed. Certificates are viable approach for restricting the amount of messages coming to the user. Architecture itself does not consider the type of object pushed on the user's device.

1

Introduction

For the past few years the number of portable devices like mobile phones, laptops and PDAs have been increasing very rapidly. People have become familiar with this development and new models of devices are introduced frequently. At the same time completely new connection technologies have emerged and different upgrades have been developed. For example, GSM networks have been upgraded with HSCSD, GPRS and EDGE technologies, WLAN technology has become very popular and Bluetooth has been mentioned more and more often. Unfortunately the portable devices and connections mean nothing unless meaningful services could be offered for the users. Too often the development is based more on the technology than for the usability of the services. In most cases the only goal for the service is the maximized profit for the developer. This paper takes another viewpoint and approaches the development through the usage scenario of some services. Possible solutions are studied and an approach for this problem is proposed. 2

Usage model: Flash N o t e

The Flash note usage model can be defined as the problem of getting localized, up-to-date information to the user's portable device according to the users personal wishes. For fluent operation, it is important that the notes pop-up on the terminal screen without user interaction. The user has to be able to define beforehand the messages he wants to receive. Examples of situations 91

92

where the user may want some localized information are presented in the following list: • Hospital/aircraft warnings, to give instructions about the usage rules • Safety instructions and emergency notes • Advertisements at shops, special prices and additional information about the products • Bingo hall to show the rolled number on user's device • Urgent messages shown when arriving at office, such as call requests • Status of home devices when arriving at home All these are situations where the user would like to get the information straight to the portable terminal without too complex procedures. The only difference between these services is the way the information is provided. In some cases the information should arrive automatically, e.g. hospital warnings. In other cases the user should be given the possibility to select the type of information received, e.g. advertising. However, in both cases the connection between the server and the portable device should be automatic without any additional acceptance by the user. If accepting automatically all the advertisements the user may be drowned by heavy amount of unwanted information. Therefore the user should also be able to decide which organizations are to be trusted and thus given the permission to send the messages. In this paper the approach for localized, automated and authenticated Flash notes is based on a wireless Bluetooth technology. The following chapters present the Bluetooth approach as well as the required parts with this technology. In this paper this architecture is referred as Flash GORDoN (Generic Object Relayed Directly on Note screen). 3

Bluetooth and Flash note

The idea behind the Flash note is the same as in the class 01 SMS (Short message service) that pop-up immediately on the users terminal screen 3 . Unfortunately SMS messages can't be refused based on their contents or source. Luckily the cost of SMS message decreases the eagerness of spamming. In case of Flash note over Bluetooth, the transmission of messages is free, which may result in extensive amount of unwanted advertisement.

93

On many of the example cases, the relevance of messages is based on the location of the terminal. Information about closing the phone when arriving to hospital is not relevant when you are at home and safety instructions near a crash site are not relevant when you are shopping miles away. Since Bluetooth wireless technology has a short operation range it ensures that user is near the area where the notification matters. Bluetooth communication is based on 2.4 GHz RF technology. The range of communication is with current specification either 10 or 100 meters. Due to this short range the services provided through Bluetooth can be easily localized. Bluetooth also allows ad hoc networking thus making the technology suitable for flash note type of services. 2 The development of Bluetooth technology has the suitable approach as the services and operation are defined by using usage scenarios. Bluetooth SIG has specified the following usage models for the Bluetooth devices • Three in a one phone • Internet Bridge • Interactive conference • Ultimate headset • Automatic synchronizer The flash note type of a service could be combined into each of these scenarios. However the automatic synchronizer is closest to the proposed approach. The automatic synchronizer is a usage model where phone, laptop and PDA devices can automatically synchronize the calendar, phone book, etc. information. The operation of this model is very close to the flash note scenario as elements are automatically transferred to the users terminal. Unfortunately this approach does not allow the use of several servers nor the groups of servers. Therefore the operation of flash notes is defined as a separate usage model. The usage models are further divided into profiles that take care of the actual definition of the operation. The profiles practically define how the protocol stack behaves with different applications. Certain profiles, e.g. generic access profile, are always needed, whereas special operations are implemented with their own profiles. On Bluetooth profile stack the Flash profile would reside on top of GOEP (generic object exchange profile). In fact flash resemble quite much of the object push profile. Both in flash and object push profile the message gets pushed to the OBEX server. The difference between

94 those is that for Flash the sender authorization is done by certificates and the incoming note is displayed immediately. From different communication protocols defined for Bluetooth, Flash GORDoN uses OBEX (Object Exchange) 5 . OBEX protocol has been adopted from IrDA (The Infrared Data Association) 4 protocols and it is targeted for exchanging simple data objects, such as messages. Since many of the existing mobile devices support IrDA communications, the OBEX protocol stack already exists on them. Thus the usage of OBEX enables also the possibility to use IrDA as the communication medium instead of Bluetooth. OBEX connection can be considered secure as it only allows simple message exchange, unlike full IP connection. OBEX is also very light protocol and do not cause much overhead for transmitting simple messages. While in this paper it is presumed that the note is plain text, this does not have to be so in all scenarios. Different types of data objects such as graphics or short mp3 audio clips can be pushed in same way. The way they are presented to the user depends highly on the user device and the implementation of the flash on it. While pushing audio messages directly to the headset would be interesting way for using flash, that is not discussed on this paper. Since there are several useful cases it is pretty clear that not all users want to get all the available information. In this paper the problem is solved by using certificates for the authentication. Certificates identify the sender and the higher authority that has granted the certificate, thus granting access permissions for either individual chips or for groups The following subchapters present how authentication and certificates could be used in the implementation of the Flash GORDoN profile. 4

Authentication

Every Bluetooth device has a unique 48 bit Bluetooth device address (BD_ADDR). This address is used for identifying the Bluetooth modules at the connection stage. When looking for other devices by using Bluetooth wireless technology, the device sends inquiry message at the baseband layer. The response coming from other devices include their BD_ADDR and this information is used for creating the connections between devices. Since BD_ADDR can be used to identify devices, it can also be used for authentication purposes for flash notes. After the connection has been created over L2CAP, the FLASH software checks the BD_ADDR of the sender and decides whether to accept the FLASH message or not. To avoid getting spammed by messages, the default policy should be rejecting the flash note

95

unless BD_ADDR is on the list of trusted flash senders. While authorizing Bluetooth enabling devices according to their B D A D D R , the list of acceptable flash note senders would eventually grow big. Another problem arises if flash message is needed to remind something urgent e.g. closing the mobile phone, when entering to the hospital. Before getting the warning message, the user would have to first add the B D A D D R of the reminder device on his mobile phone. Since there can be multiple entrances to the hospital, this means also multiple reminder devices and thus lots of work from users point of view to add all the entrances. The same routine to add all entrances happens when the user enter to the other hospital. This clearly makes the flash notes useless for this kind of situations unless better ways of authentication is developed.

Figure 1. Certificate issuer tree

Instead of using plain BD_ADDR to decide whether to accept or to reject the note a certificates similar to those used at PKI (public-key infrastructure) could be used to authenticate the flash sender. The certificates for devices are issued by certificate authority (CA). An emergency certificate could be authorized to the places like hospitals and police vehicles, and issuer for these could be goverment. The certificate issuing tree about this is shown in Figure 1. Different supermarket coalitions could have their own Certificate Authority (CA) that issues the certificates to different markets for advertising. This way a customer can approve advertisements from either one market or from the whole coalition. Approving the CA of McDonalds would result McDonalds special offers to come fluently on the screen of PDA all over the world without the need of adding every single burger place to the authenticated list. If the user wants only the special price information from local McDonalds he could approve only the local certificate, but leave global CA signed certificates disapproved.

96 5

Certificates

Certificates has been used for authenticating someones public-key and are essential part of public-key infrastructure (PKI) such as x.509 7 8 . They can be used also to authenticate group of devices instead of only one. For flash note type of application PKI systems hold more information than necessary. The only required information is the organization who has issued the certificate and to whom it has been issued. The certificate must depend on the B D A D D R of the certificate holder thus B D A D D R is the identification of the certificate holder and all certificates are unique. When certificate arrives to the receiver, receiver can compare the certificates owner id to the sender device's BD_ADDR to make sure the certificate is valid. For these purposes a simplified version of x.509 can be used where BD_ADDR of the device is used as the distinct name. To create a certificate for a device, CA uses public-key cryptography, such as RSA 6 . CA takes the B D A D D R of the device it wants to create the certificate for and encrypts it by using the secret key. The result of encryption is combined with the identification information about the CA. To make certificate structure consistent, the CA should also be Bluetooth device with it's own BD ADDR. CA

BD_ADDR[CA] E(BD_ADDR[CA],PKEY[CA])

PKEYfCA]

__.__----""" Dl

Cert[CA]

E[s,CA](BD_ADDR[Dl],PKEY[Dl])

PKEY[D1]

""""" D2

Cert[Dl]

E[s,Dl](BD_ADDR[D2],PKEY[D2])

t Certificate holder field

PKEY[D2]

t Holder public key field

Authenticator certificate field Figure 2. Certificate tree

The certificates should be saved on the users device so that the certificate is paired with the B D A D D R it belongs to. When flash note sender presents certificate, the user device first checks, if the sender device is in the trusted

97

listed. If not, then the device goes the certificate tree towards the root, until it finds either a trusted certificate or trusted B D A D D R in the authenticator certificate field. When finding trusted certificate, user device decrypts the certificate holder field. Prom there he gets B D A D D R and the public-key of that certificate holder. 1: If the certificate holder is the device that sent the flash, user device compares the B D A D D R of sender device and the one it got from certificate. If the addresses match, the certificate is valid and the flash message can be shown. 2: If the certificate holder is device that has issued the certificate that needs to be verified, user device compares the Public-key decrypted from the certificate holder field to the one at holder's public-key field. If the public-keys match, the certificate holder is valid issuer, and the issued certificate can be verified with the public-key. There exits a possibility to fool the system, by creating Bluetooth device that sends false B D A D D R and then copies certificate given to that B D A D D R . This can be prevented by sending extra message in challengeresponse manner to verify that certificate holder actually knows the secret key matching to the public key on certificate. We feel this unnecessary since the gain for forfeiting the certificate compared to the trouble is not worth it. 6

Binding the parts together

In order to be able to receive flash notes, the user's device has to enable OBEX connections over Bluetooth. Flash note application on user's device works as the OBEX server. It allows other Bluetooth devices to connect to it for pushing flash notes. The flash note service needs to be registered in the device's service discovery database, in order to enable other devices to find the service. Once the server is running background of the device, it can be contacted by devices that send the flash notes. The sender goes through the normal Bluetooth connection procedures. (Figure: 3) It forms a basic connection to the user device and requests the available services. If the user device is running a flash server, sender pushes its certification to the server. The user device then checks the certificate's authenticity and either accepts or rejects flash connection depending on the users definitions. If user device accepts the connection, it sends ok message to the sender. After getting acknowledgement from the user device, the sender pushes the flash note to the server. When server gets the message, it pops it to the screen of the device. Another approach would be so that sender pushes both certificate and

98 User

Server

Form L2CAP connection

SDP _ServiceSearchRequest(Flash)

If user can understand FLASF

Setup RFCOMM link for OBEX Open Obex session id+ certificate

Accept/reject certificate Send FLASH

Figure 3. Message Sequence Chart of Flash GORDoN

the flash message simultaneously. The server would show the message only if the certificate is on its list of authorized senders. After that server replies with either ok or fail message, depending whether it accepted the flash note or not. This way if the sender message was rejected, it knows this and do not waste the bandwidth by sending messages that will not be shown.

7

Conclusions

In this paper Flash GORDoN architecture has been presented. It allows customers to define the sources of incoming messages according to the personal preferences. It utilizes Bluetooth as the communication medium and OBEX protocol is applied for transmitting the messages from service provider to the user. Authentication hierarchy is implemented with certificates. This approach allows bundling the potential note providers together into groups of trusted sources and forms a viable approach for restricting the amount of messages coming to the user in the jungle of advertisements.

99 References 1. Bluetooth SIG, Bluetooth specification 1.0B. 1999 2. Bray, Jennifer. Sturman, Charles F. Bluetooth: Connect without Cables, Prentice Hall 2001 3. ETSI, GSM 03.40 standard 4. IrDA, IrDA homepages, http://www.irda.org/ 5. IrDA, OBEX Specification, 6. Menezes,Alfred J. van Oorschot Paul C. Vanstone ScottA. Handbook of Applied Cryptography, CRC Press 1997 7. Housley R. Ford W. Polk, W. Solo D. RFC 2549 ,The Internet Society 1999 8. Schneier B. Applied Cryptography, John Wiley & Sons 1996

PERFORMANCE ENHANCEMENTS TO THE IEEE 802.11B STANDARD MATTHEW B. SHOEMAKE Texas Instruments, Inc. 12500 Tl Boulevard, Dallas TX 75245, USA E-mail: [email protected] Since its adoption in September of 1999, the IEEE 802.11b standard has enjoyed great commercial success. This success has resulted in further standardization efforts to enhance and extend the performance of IEEE 802.11 based networks. In March of 2000, the IEEE 802.11 Working Group approved the formation of a Study Group to investigate the feasibility of extending the IEEE 802.11b standard to data rates in excess of 20Mbps. After initial investigation in mid-2000, the committee determined that such extension was practical and the Task Group G of IEEE 802.11 (IEEE 802.1 lg) was formed to standardize such an extension to the 802.1 lb specification. The IEEE 802.1 lg committee has been presented with multiple types of performance enhancements including data rate enhancements, coding enhancements and enhancements to the preamble and channel estimation capabilities of the preamble of each packet. This paper contains detailed descriptions of each of the major enhancements that have been proposed. The paper also outlines recent regulatory changes that affect IEEE 802.11 g.

1

Introduction

This paper illuminates proposals to the IEEE 802. llg standards committee and issues related to this standardization. Some attention is given in to recent regulatory changes that have allowed such enhancement and extension in the 2.4GHz band. Key performance metrics such as throughput and packets per second are evaluated and plotted for the Barker, CCK and PBCC modes of the IEEE 802.11b standard as well as for the newly proposed PBCC and CCK-OFDM modes. Special attention is given to the throughput for when the payload is a small number of bytes, since this case is often not focused on but accounts for a significant percentage of traffic. Range versus rate curves are also shown for specific modulation schemes being considered in IEEE 802.llg. 2

Regulatory Changes

The FCC has recently announced regulatory decision-making processes in the United States that will allow for the use of non-spread spectrum devices in the 2.4GHz ISM band. It is the view of the author that this is a positive development, as it will align the regulations of the United States with those of other countries that do not require spread spectrum modulation to obtain approval to operate in the band. In addition, the official measurement techniques used to decide if a communications device qualified as spread spectrum were not accurate. Non-

100

101

spread spectrum modulations such those contained in the 802.11b standard, passed the FCC test and were certified as being spread spectrum devices and therefore allowed to operated in the 2.4GHz ISM band. Ironically, the fact that these nonspread spectrum devices were allowing to operate in the band resulted in a tremendous increase in the utilization of the band, a public resource. The lifting of restrictions requiring spread spectrum modulation will allow for higher spectral efficiencies while generating no more interference than devices that are already allowed to operate in the band. Specifically, the FCC has issues a Notice of Further Rule Making for the 2.4GHz ISM band. In doing so, the FCC indicated that until the rules making was completed, it will allow waivers for some technologies including PBCC™ and OFDM modulations. These waivers will allow IEEE 802. llg to adopt a standard that can be certified in the United States and worldwide. 3

Coding Enhancements

The IEEE 802.1 lb standard is based on three types of modulations, namely Barker modulation, Complementary Code Keying (CCK) and Packet Binary Convolutional Codes (PBCC). Barker modulation is a true spread spectrum system with a spreading factor [4] of 22 when operating at 1 Mbps and a spreading ratio of 11 when operating at 2 Mbps. CCK is based on a block coding scheme, and PBCC is a convolutional coding scheme. Both single tone and multi-tone extensions to IEEE 802.11b have been proposed. The leading single tone candidate is the PBCC™ proposal [3]. The PBCC proposal focuses on extending the 5.5 and 11Mbps PBCC modes in the 802.11b standard with convolutional coding modes at rates of 16.5, 22, and 33 Mbps. There are two underlying convolutional codes used in PBCC modulations. A 64state code is used for rates of 5.5 and 11Mbps and a 256-state code is used for 22 and 33Mbps. The spectral mask requirement of DEEE 802.1 lg is mandated by its IEEE Standards Association project authorization to be the same as that of the IEEE 802.11b standard. The spectral mask of the IEEE 802.11b standard is quite liberal given its symbol rate of 1 lMsps. A square root raised cosine pulse shaping filter with bandwidth expansion factor as high as 0.8 can be used and still meet the mask. The 33 Mbps mode of PBCC is obtained by operating the 22Mbps mode at a symbol rate of 16.5Msps.

102 Tablet. PBCCModulations

Symbol Rate (Msps) 11 11 11 16.5

BCC States

Constellation

Code Rate

64 64 256 256

BPSK QPSK 8-PSK 8-PSK

Vi

Coding Gain (dB) 5.9 5.9 8.1 8.1

1 1 2/3

Data Rate (Mbps) 5.5 11 22 33

The PBCC packet structure is identical to that of IEEE 802.11b. Each packet consists of a preamble, a header and a physical layer convergence protocol service data unit (PSDU). The preamble is deterministic and used for packet detection, synchronization and channel estimation. The header contains information about the PSDU such as modulation type and number of bytes of data. The PSDU contains only encoded data. The data is encoded using Barker, CCK or PBCC modulation, where Barker is a true direct sequence spread spectrum scheme, CCK is a block coding scheme, and PBCC a convolutional coding scheme. The packet structure is shown below in Figure 1. Preamble 72 or 144 bits

Header 48 bits

PSDU ^

1,2 Mbps Barker 5.5,11 CCK 5.5,11,22,33 PBCC

Figure 1. IEEE 802.1 lb and PBCC Packet Stnicture

The only multi-tone scheme proposed as an extension to IEEE 802.11b is known as CCK-OFDM [1,2]. This proposed extension to IEEE 802.11b results in hybrid single-tone and multi-tone packets. The preamble and header for a CCKOFDM packet is the same as that of IEEE 802.1 lb. This is actually a requirement to be backward compatible with legacy IEEE 802.11b devices. However, the packet structure differs significantly from IEEE 802.11b in that the PSDU. The PSDU contains a deterministic synchronization field, a signal field, OFDM data symbols and a signal extension. The CCK-OFDM packet structure is shown below in Figure 2.

103 Preamble 72 bits

Header 48 bits

^

^

OFDM SYNC 8 us

PSDU

^ OFDM OFDM Symbols Signal Extension Signal 6 us

6,9,12,18, 24,36,48,54 Mbps Figure 2. CCK-OFDM Packet Structure

The CCK-OFDM proposal draws from the IEEE 802.11a standard, an OFDM based standard targeted at the 5GHz U-NII band. The IEEE 802.11a standard is based on a symbol rate of 20Msps, while the IEEE 802.11b standard is based on a symbol rate of 11Mbps. The CCK-OFDM proposal leaves the symbol rate 11 Msps through the preamble and header and changes the rate to 20 Msps for the PSDU. This change of symbol rate generates the need for a second synchronization and signal field prior to transmission of the data. In addition, since CCK-OFDM decoder has more latency than Barker, CCK or PBCC, an additional signal extension that conveys no information and consists of random data is transmitted to artificially extend the length of the packet. If the packet were not artificially extended, a typical CCK-OFDM decoder could not decode the PSDU in less than the lOus limit imposed by the IEEE 802.11 MAC layer. The data rate for the OFDM symbols can be 6, 9, 12, 24, 36, 48 or 54 Mbps. The OFDM symbols are coded using a 64-state binary convolutional code. 4

Rate Enhancements

The instantaneous data rate achieved in the payload portion of the packets is greater than the sustained throughput that can be achieved. In fact, the instantaneous data rate can be quite deceiving. For example, IEEE 802.11b networks can achieve an instantaneous data rate of 11Mbps, however, the sustained throughput due to MAC and PHY is much lower. IEEE 802.1 lb systems using the Distributed Coordination Function (DCF) of the IEEE 802.11 MAC and the long,

104

192 us preamble, can only maintain a sustained throughput of 6.5 Mbps for 1500 byte packets. At a packet size of 64 bytes a throughput of only 0.6 Mbps can be maintained. Overhead in the networks comes from MAC layer overheads and PHY layer overheads. Calculation of die maximum sustainable throughput of an IEEE 802.11 network under DCF operation can be calculated as follows. Rs = 8xNxR,jiN) where N is the number of bytes per packet, Rp(N) is the number of data packets per second and Rs is the sustainable throughput in Mbps. ACK

Data SIFS = 10 us M

^

M

fc

DIFS= 50 us

•*

fc-

Data

Backoff

TP,DCF(N)

w

Figure 3. - DCF Average Transmission Time Calculation

The number of packets per second can be calculated as the reciprocal of the average amount of time to transmit a packet, TPDCF(N). The average backoff when there are no collisions is the mean of the number of slots backed off times the duration of a slot. The average number of slots backed of is 15.5, and the slot duration is 10 us, thus the average backoff is 155 us. Thus Tp,DCf(N>

R> TpH, Tpjjy Overhead)

=

2TPH + 2TPHYOverhead + 8N/R + 10 + 112/R + 50 + 155 us = 2TPH + 2TPHY overhead + 8N/R + 112/R + 215 US

where TPH is the time for the preamble and header, TPm overhead is m e physical layer overhead and R is the data rate of the payload. The time for the preamble and header currently take on one of two values, namely 192 or 96 us. The time spent on other physical layer (PHY) overhead varies depending on the coding scheme. The rates, R, that can be achieved also vary depending on the coding scheme.

105 Table 2. PHY Overheads

Modulation Scheme Barker CCK CCKOFDM PBCC PBCC PBCC PBCC

R (Mbps) 1 5.5, 11 6, 9, 12, 18, 24, 36, 48, 54 5.5 11 22 33

TpHY Overhead

(us) 0 0 18

1.4545 0.7273 0.3636 0.2424

Given the values above, TPDCF(N, R, TPH, TPHYOverhead) can be calculated and plotted against N. Figure 4 below shows TpDCF(N, R, 96 us, TPHY overhead) m packets per second versus N for IEEE 802.11b including the optional PBCC modes plus the new 22 Mbps mode that has already been implemented by Texas Instruments.

-•—BarkerlMbps ••—Barker 2 Mbps -A—CCK 55 Mbps •K— CCK 11Mbps • * — PBCC 5.5Mbps -•—PBCCTIMbps -I—PBCC 22Mbps

N (bytes) Figure 4. TP,DCF(N, R, 96 us, Tpmo^^ad) versus N for Barker, CCK and PBCC

The above figure shows that the throughput for CCK and PBCC at 5.5 and 11 Mbps are very similar despite the very small additional overhead for PBCC this small amount of overhead is due to zero forcing the state machine at the end of the data packet. Likewise, it can be seen that the 22 Mbps PBCC scheme achieves well over 12 Mbps of sustained throughput under DCF operation.

106

The sustainable throughputs in Mbps for the PBCC modes of IEEE 802.11b are plotted along with those of the new PBCC modes and the newly proposed CCK-OFDM modes in Figure 5 below. It can be seen that under DCF operation, the sustained throughputs are still significantly less than the instantaneous throughputs. For 1500 byte packets, 33 Mbps PBCC reaches a sustained throughput of approximately 16 Mbps, a ratio of just less than one-half. For 1500 byte packets, 54 Mbps CCK-OFDM achieves a throughput of 18 Mbps or about 35% of the instantaneous rate.

PBCC 5.5 Mbps PBCC 11 Mbps PBCC 22 Mbps PBCC 33 Mbps CCK-OFDM 6 Mbps CCK-OFDM 9 Mbps CCK-OFDM 12 Mbps CCK-OFDM 18 Mbps CCK-OFDM 24 Mbps CCK-OFDM 36 Mbps CCK-OFDM 48 Mbps CCK-OFDM 54 Mbps

Figure 5. R, versus N for PBCC and CCK-OFDM

5

Preamble Enhancements

The previous section shows that there is significant overhead reduction that can be attacked in trying to increase the performance of IEEE 802.11b networks. Preamble enhancement is a key area that the IEEE 802.llg body can focus on. The two legacy preambles of 96 us and 192 us are significantly longer than required to estimate the multi-path experienced by 802.11-based WLAN's in the 2.4GHz band. Preambles have been designed that provide significant improvements over these systems. The structure of such a preamble is shown in Figure 6 below.

107

TIMING SYNC FRAME SYNC CHANNELEST HEADER 96 symbols @ 1 lMsp|64symbols @ HMsp 68 symbols @ UMs448 BITS @ 5.5Mbps

PSDU

Figure 6. Enhanced Preamble Structure

This preamble reduces overhead by approximately 58 us per packet with respect to the IEEE 802.1 lb short preamble. When used with PBCC, there is very little overhead at the physical layer that is left. The enhanced preamble structure above was designed to operate and allow channel estimation and detection at SNR's as low as 2.5 dB. This allows operation down to 5.5 Mbps PBCC, but it precludes operating 1 Mbps Barker all the way down to its threshold. 6

Range versus Rate

Ultimately the true measure of the performance of a wireless LAN system can be obtained in a view key parameters such as delay, jitter, throughput and range. These parameters may be divided into those that are impacted most at the physical layer and the medium access layer. At the physical layer, die range versus throughput of a system is key. The IEEE 802.llg committee has used a slightly more forgiving metric to measure maximum sustainable throughput. In fact, the metric that is used by die committee is currently unachievable, because MAC overhead is currently not as low as it assumes. The logic of the metric though, it that it assumes a very low MAC overhead, which forces the IEEE 802.1 lg committee, which is a PHY committee, to focus on reducing PHY overhead. The metric used by die commit is capture by the follow equation Tp,llg(N,

2TPH + 2TpHYOverhead

R, TPH,

+ 8N/R

TPHY Overhead)

+ 10 + 112/R

2Tp„ + 2TPHy overhead + 8N/R

+ 112/R

=

+ 50 US = + 60 US

Using the above metric, the following range rate points have been obtained as shown in Figure 7. The performance of the IEEE 802.11a standard, if it where operating in the 2.4 GHz ISM band, is also shown in the figure. Note the significant difference between CCK-OFDM range rate performance and mat of

108

IEEE 802.11a, due to the additional overhead incurred by the hybrid single tone, multi-tome CCK-OFDM system.

Range vs. RatetorPER = 10" 35

i

54

j

I

"\48 30V

-<— -©•• -•

-

CCK PBCC 802.11a-OFDM CCK-OFDM

-

' J36 25-

«-20 • a.

f

54

tu

24;

«~

s

K*1Sr

"'•••>

1 8 ^

---...

22 >; ~ - - " 12

""13...;

24 f ""-

t8-'r^-"'

11

- ^

---^12 55

i

°40

60

i

80 100 120 Range normalized with respect to CCK11

5.5 i

140

160

Figure 7. - Range versus Rate

7

Acknowledgements

I thank Anuj Batra and Chris Heegard for their continued effort to aid in technical analysis of the IEEE 802.11b and proposed IEEE 802.11g standards. References 1. Baldwin Keith, Halford Steve, Webster Mark and Zyren Jim, CCK-OFDM Proposed Normative Text, doc. IEEE 802.11/01-437. 2. Halford Steve and Webster Mark, Overview of OFDM for High Rate Extension, doc. IEEE 802.11-01/389.

109 3. Heegard Chris, Rossin Eric, Shoemake Matthew B., Coffey Sean and Batra Anuj, Texas Instruments Proposal for IEEE 802.llg High Rate Standard, doc. IEEE 802.11-00/384. 4. Massey James L., Towards an information theory of spread-spectrum systems. In Code Division Multiple Access Communications, ed. By S. G. Glisic and P. A. Leppanen (Kluweer, Boston, 1995) pp. 29-46. 5. Shoemake Matthew B. and Batra Anuj, Range versus Rate Comparison of Remaining IEEE 802.llg Proposals: PBCC and CCK-OFDM, doc. IEEE 802.11-01/286.

Part 5 Network Security

A USER AUTHENTICATION SYSTEM FOR SECURE WIRELESS COMMUNICATION NARIYOSHI YAMAI Computer Center, Okayama University 3-1-1, Tsushima-naka, Okayama 700-8530, Japan E-mail: [email protected] HAYATO ISfflBASHI Media Center, Osaka City University 3-3-138 Sugimoto, Sumiyoshi-ku, Osaka 558-8585, JAPAN E-mail: [email protected] HIDEO MORISHITA AND TOSfflAKI MORI Stella Craft, Inc. 1-4-20, Sonemki-shinchi, Kita-ku, Osaka 530-0002, JAPAN E-mail: {manmos,toshiaki}@stellar.co.jp KOTA ABE AND TOSHIO MATSUURA Media Center, Osaka City University 3-3-138 Sugimoto, Sumiyoshi-ku, Osaka 558-8585, JAPAN E-mail: {k-abe,matsuura}@media.osaka-cu.ac.jp In wireless LAN environments, access control mechanisms are required so that only authorized users are allowed to access the network. However, existing IEEE 802.11b equipments only provide ESSID, MAC address filtering and WEP, which are too weak to ward off malicious attacks. To prevent malicious users from unauthorized access to the wireless LANs, we have developed an access control system of IEEE 802.lib wireless LAN with a user authentication mechanism. This system provides prevention functions not only from network access by unauthenticated users, but also from packet spoofing by authenticated users. In addition, access control function per user and logging function of users' activities are also provided in the system.

1

Introduction

As mobile PCs and wireless LAN equipments proliferate, it is common for people to access to networks from their mobile PCs via wireless LAN. Especially, IEEE 802.11b [1] equipments are commonly used and many organizations provide wireless LAN environments with these equipments. In such environments, since anyone can access to the wireless LAN, access control mechanisms are required so that only authorized users are allowed to access

113

114 the network. Additional mechanisms are needed to record who, when, and from where the network accessed, in order to prevent network abuse. As access control mechanisms on IEEE 802.11b, ESSID (Extended Service Set Identifier), MAC (Media Access Control) address filtering and WEP (Wired Equivalent Privacy) are provided. ESSID and MAC address are, however, not so safe since they are included in wireless packets and malicious users can obtain and use them by monitoring wireless packets. WEP uses a secret key to encrypt/decrypt wireless packets. Although this key is not included in wireless packets, it is shared by all users of the wireless LAN. Therefore, the users cannot be identified with this key. In addition, it is difficult to keep away the key from malicious users if many users share the key. On the other hand, IEEE 802. IX [2] provides an individual key for each user. Therefore, malicious users cannot gain access to the wireless LAN. However, IEEE 802. IX is under development and is not widely used. There still exist many wireless LAN equipments without IEEE 802.IX functions. To prevent malicious users from gaining unauthorized access to the wireless LANs, we have developed an access control system of IEEE 802.11b wireless LAN with a user authentication mechanism. This system provides (1) prevention function from unauthorized access, (2) access control function per user, and (3) encryption function of wireless packets. Function (1) includes two subfunctions: (la) prevention function from network access by unauthenticated users and (lb) prevention function from packet spoofing by authenticated users. Although function (3) is implemented with the common key encryption scheme, our system does not require key distribution in advance. In the following sections, we describe the design and implementation of the system. 2

Design of the System

Before we develop the system proposed in this paper, we developed a user authentication system for wired LAN sockets, called LANA (LAN Authentication) system [3]. In addition to user authentication function that authorizes the users of LAN sockets, LANA system provides the following functions: an access control function that allows only authenticated users to send packets to the backbone network, and a protection function that prevents IP and/or MAC address spoofing, thus preventing malicious users from abusing the network. To realize these functions, LANA system uses L2 switches with IEEE 802.1Q VLAN tagging features [4] and assigns a unique VLAN ID to each physical port of the L2 switches. Since VLAN IDs are inserted to all the frames by the LAN switches, users cannot control the IDs. Therefore, a user's PC can be uniquely identified by VLAN ID assigned to each switch port. LANA system uses the 3-tuple of its IP address, MAC address, and VLAN ID to protect against unauthorized network access.

115

PC

r V.

Backbone Network

1. Connects to the wireless LAN

A \ J

4. Authenticates the PC

5. Forwards the packets PC

Figure 1 Outline of the system

However, on wireless LAN environment, since a user's PC is connected to a logical channel, rather than a physical port, of the access point, it is difficult to identify users' PCs certainly if packet spoofing is possible. In this paper, in order to identify a user's PC, we introduce a uniquely identified logical connection between the PC and the gateway of the wireless LAN, through which all the packets sent from or to the user's PC flow. Since this connection is established after user authentication, only authorized users can access to the wireless LAN. Moreover, all the packets flowing on this connection are encrypted so that malicious users cannot hijack the connection. Figure 1 shows the outhne of the system. This system consists of client PCs, an access point (AP), an authentication system, and an IP address server. As shown in this figure, client PCs are connected to the authentication system via the AP. The authentication system provides packet forwarding function between client PCs and the backbone network as well as user authentication function. It also provides userspecific access control function. That is, this function controls the services available to each user. The system runs as follows: 1. Users connect their PC to the wireless LAN. 2. Each PC obtains an IP address from the IP address server. 3. Each PC establishes a secure connection to the authentication system using some encryption technique such as SSL (Secure Socket Layer). 4. The authentication system authenticates the PC using some personal identification information such as a user name and a password.

116

Backbone Network

Figure 2 Outline of the system

5.

After the PC has been authenticated, the authentication system allows the PC to communicate with the backbone network.

3

Implementation of the System

3.1

System Configuration

We have implemented a prototype system called WILL (Wireless LANA) system. The system configuration is shown in Figure 2. This system consists of users' PCs, an access point, a LANA filter, a LANA server, a DHCP server, and a RADIUS server.

117

For make the LANA system available on the wireless environment, we introduce a PPP connection over an SSL session for user authentication and encryption. For this reason, a PPP server and an SSL server are installed in the LANA filter as its components. In addition, a special purpose program dealing with this connection, called a WILL client, is introduced on the users' PCs. In the following subsections, each component of the WILL system is described in detail. 3.1.1

LANA filter

Filter modules: The LANA filter forwards packets among users' PCs, the servers on the WILL system, and the backbone network, with the following filtering function: 1. It always forwards DHCP packets between users' PCs and the DHCP server. (Figure 2(a)) 2. It always forwards packets on the SSL session between users' PCs and the SSL server on the LANA filter. (Figure 2(b)) 3. It always forwards packets via the SSL session between users' PC and the LANA server. (Figure 2(c)) 4. After the PC has been authenticated successfully, it forwards packets via the PPP connection between users' PC and the backbone network, along with source address checking and user-specific packet filtering. (Figure 2(d)) Note that the functions 1-3 are performed on the incoming filter module (Filter(in) of Figure 2), whereas the function 4 is performed on the outgoing filter module (Filter(out) of Figure 2). SSL server module: This module waits for a connection originating from a user's PC, and then it establishes an SSL session and forwards packets between the PC and the LANA server. After the LANA server authenticates the PC, this module forwards packets on the SSL session between the PC and the PPP server. PPP server module: This module establishes a PPP connection on the SSL session to the PC (WILL client). LANA server: The LANA server authenticates the user of the WILL client and manages user-specific filtering setting if necessary, by keeping track of users' activities such as login and logout. Instead of IEEE 802.1Q VLAN ID on the former LANA system, the LANA server on the WILL system associates an authenticated PC with PPP connection ID.

118 3.1.2

WILL client

The WILL client establishes an SSL session to the LANA filter on start up. Then, according to die request from the LANA server, it sends the user identification information. After authentication is done successfully, it establishes a PPP connection on the SSL session, and makes all the packets flow through the connection. 3.1.3

DHCP server

The DHCP server assigns an appropriate IP address to the PC. Any DHCP server will work and we use one of free servers, ISC DHCP server [5]. 3.1.4

RADIUS server

The RADIUS server performs user autiientication and accounting. Any RADIUS server will work too and we use one of free servers, DTC RADIUS [6]. 3.1.5

Access point

The access point is one of IEEE 802.11b equipments and acts as a MAC bridge between the wireless LAN and the wired LAN. 3.2

Connection sequence

An connection sequence in detail is depicted in Figure 3 and is described as follows: 1. Requesting an IP address to the DHCP server A user's PC requests the DHCP server to assign an IP address to the PC. This address is only used for establishing an SSL session. Therefore, the packet filter allows all the packets sent from this address only to go to the SSL port of the LANA filter. 2. Establishing an SSL session The WILL client establishes an SSL session to the SSL server on the LANA filter, and then connects to the LANA server with this session. The LANA server requests the WILL client to send back user identification information, namely the user name and the password. This information is then forwarded to the LANA server, and the LANA server authenticates the user using the RADIUS server. If this authentication is passed successfully, then the LANA server sets up an appropriate filtering function on the LANA filter and turns on the RADIUS accounting. Otherwise, this session is closed. 3. Establishing a PPP connection The WILL client establishes a connection to the PPP server over the SSL session. 4. Obtaining another IP address from the PPP server

119

Backbone Network * User's PC

j 3. Establishes a PPP connection over the SSL session 4. Obtains another IP addresr " ~ - - — - - - - - 5. Updates the routing table Figure 3 Connection sequence

After establishing a connection to the PPP server, the WILL client obtains another IP address from the PPP server using IPCP [7]. This address is used for communication with the outside of the system. Updating the routing table on the user's PC The user's PC updates its routing table to change the default routing information to the PPP server. Accordingly, the PC can communicate with the backbone network via the secure PPP connection, with an appropriate filtering function. 6. Terminating the connection If me WILL client terminates the PPP connection, or if no packets are detected on the connection for a specific period,tiienthe LANA server turns off the RADIUS accounting, closes the SSL session (and the PPP connection if necessary), and removes the corresponding filtering function on the LANA filter. 4

Conclusions

In this paper, we have proposed a user authentication system that provides secure communication on IEEE 802.11b wireless network environment. With this system, unauthorized network access is not allowed even if IP and/or MAC addresses are spoofed. We also have shown an implementation of the system called WILL system. Further works include performance evaluation of the WILL system, developing roaming function between WILL systems and between WILL and (wired) LANA systems, and so on.

120 References 1.

2.

3.

4. 5. 6. 7.

IEEE, 802.1 lb-1999 Supplement to 802.11-1999, Wireless LAN MAC and PHY specifications: Higher speed Physical Layer (PHY) extension in the 2.4 GHz band, IEEE (1999). IEEE, P802.1X, DIO, January 2001 DS5886 P802.1X Draft Standards for Local and Metropolitan Area Networks: Standard for Port based Network Access Control, IEEE (2001). Ishibashi H., Yamai N., Abe K. and Matsuura T., A Protection Method against Unauthorized Access and Address Spooring for Open Network Access Systems, Proceedings of 2001 IEEE Pacific Rim Conference on Communications, Computers and Signal Processing (PACRIM '01), 1 (2001) pp.10-13. IEEE, 802.1Q-1998 IEEE Standards for Local and Metropolitan Area Networks: Virtual Bridge Local Area Networks, IEEE (1998). Internet Software Consortium, ISC DHCP, http://www.isc.org/dhcp.html. Digital Technology Inc., DTC Radius 2.03, http://www.dtc.co.jp/Radius2.0A McGregor G., The PPP Internet Protocol Control Protocol (IPCP), RFC 1332 (1992).

DESIGN, IMPLEMENTATION, AND EVALUATION OF BLUETOOTH SECURITY

JUN-ZHAO SUN, DOUGLAS HOWIE, ANTTIKOIVISTO, AND JAAKKO SAUVOLA MediaTeam, Machine Vision and Media Processing Unit, Infotech Oulu, P.O.Box 4500, FIN-90014 University of Oulu, Finland E-mail: jundiao.sun/douglas.howie/antti.koivisto/jaakko.sauvola @ee.oulu.fi Bluetooth is a recently proposed protocol for local wireless communication and has become a de facto standard for short-range ad hoc radio connections. Security concern is one of the most important problems delaying the mass adoption of Bluetooth. This article provides a study on the security issues behind the Bluetooth standard. After a overview of the general Bluetooth protocol, a security framework is introduced for the description of the Bluetooth security layout. Then both link-level and service-level security schemes are discussed in detail on the basis of the framework. Some weaknesses of the Bluetooth security strategies are analysed, together with potential risks and possible attacks against the vulnerabilities. Corresponding countermeasures are also proposed in order to improve the Bluetooth security.

1

Introduction

Bluetooth, a new technology named after the 10th century Danish king Harald Bluetooth, is a recently proposed standard for local wireless, communication and is becoming hotter and hotter a topic. The primary design goal of Bluetooth is a cable replacement protocol for wireless connectivity. Now it has extended to include the application scenarios of both voice/data access points and personal ad hoc networks. A diverse set of wired and wireless devices are Bluetooth connectable, including office appliances, e.g. desktop PCs, printers, projectors, laptops, and PDAs; communication appliances, e.g. speakers, handsets, pagers, and mobile phones; home appliances, e.g. DVD players, digital cameras, cooking ovens, washing machines, refrigerators, and thermostats. Bluetooth is suitable for a wide range of applications, e.g. wireless office and meeting room, smart home and vehicle, intelligent parking, electrical paying and banking. Bluetooth is a de facto standard for ubiquitous devices to achieve the pervasive connectivity by low-power, short-range, low-cost embedded radio. The normal transmitting power is lmW (OdBm) and the option islOOmW (-30 to +20dBm). The normal range is 10m and the optional one 100m. Power consumption is from 20mA to 30mA on different operating states. The cost of a single-chip Bluetooth solution expects to be around $5 per device. Bluetooth adopts a master-slave architecture to form an ad hoc wireless network named piconet. A master in a piconet may communicate with up to seven active slaves. Several connected piconets can further form a scatternet. Bluetooth was developed by Bluetooth Special Interest Group (SIG) formed in May 1998. The founding members included Ericsson, Intel, IBM, Nokia, and

121

122

Toshiba. Since then, almost all of the biggest telecommunications companies have joined the Bluetooth SIG, e.g. 3Com, Lucent, Microsoft, Motorola, etc. Bluetooth specification is a free open standard and the latest version 1.1 was approved in February 2001 [1, 2]. The Wireless Personal Area Network (WPAN) standard, developed by the IEEE 802.15 Working Group [3], is based on the Bluetooth. Security is always one of the most important issues to any communication technique. In a wireless scenario, such as the Bluetooth, this problem becomes more severe for the totally open-air transmission. The aim of this paper is to provide a study on Bluetooth security, including an overview of the fundamentals, an introduction of implementation issues, and an analysis of potential vulnerabilities. The rest of this paper is organized as follows. First as the background, we briefly review the Bluetooth protocol. Next, a security framework of the Bluetooth is introduced to figure out a layout of Bluetooth security. Then we examine the Bluetooth security in two parts according to the framework proposed, including the build-in link-level Bluetooth security schemes as the principle part, and the servicelevel Bluetooth security architecture as the practice part, respectively. Specifically, the weaknesses of the Bluetooth security are discussed in detail, leading to a comprehensive summary of the Bluetooth security problems, together with some possible countermeasures that we propose. Finally we sum up the discoveries and make the conclusions. 2

Bluetooth Protocol Overview

Detailed information on the Bluetooth standard can be found in part A to E of [1] and [4, 5]. Figure 1 illustrates the Bluetooth protocol stack, which can be divided into four layers according to their purpose, in the following way: 1. Bluetooth Core Protocols, including Baseband, LMP, L2CAP, and SDP, comprise exclusively Bluetooth-specific protocols developed by the Bluetooth SIG that are required by most of the Bluetooth devices. 2. Cable Replacement Protocol, i.e. RFCOMM protocol, is based on the ETSI TS 07.10 that emulate serial line control and data signals over Bluetooth Baseband to provide transport capabilities for upper level services. 3. Telephony Control Protocols, including TCS Binary and AT-commands, are used to define the call control signalling, mobility management procedures, and multiple usage models for the Bluetooth devices to establish the speech and data calls and provide FAX and modem services. 4. Adopted Protocols, including PPP, UDP/TCP/IP, WAP, WAE, etc. Due to the open nature of the Bluetooth specification, additional protocols (e.g., HTTP, FTP, etc.) can be accommodated in an interoperable fashion. 5. Host Controller Interface (HO), i.e. the boundary between hardware and software, provides a uniform command interface to access capabilities of hardware, e.g. Baseband controller, link manager, control and event registers.

123

Figure 1. Bluetooth protocol stack.

The layers of Cable Replacement, Telephony Control, and Adopted Protocols form the application-oriented protocols that enable applications to run over the Bluetooth core protocols. Not all applications make use of all the protocols shown in Figure 1. Instead, applications run over one or more vertical slices of this protocol stack. In other words, applications may run over different protocol stacks. Nevertheless, each one of these different protocol stacks uses a common Bluetooth data link and physical layer, i.e. Bluetooth core protocols, including: •

•

•

•

Baseband. Based on the physical radio link, the Baseband can form the piconet between Bluetooth units and decide the roles of master and slave in the piconet. The Baseband provides physical links of both Synchronous ConnectionOriented (SCO) and Asynchronous Connectionless (ACL) to support the transmission of data and/or audio with corresponding packets. Other functions include error correction, link management and control, audio transmission, etc. Link Manager Protocol (LMP). The Bluetooth protocol LMP is responsible for link set-up between Bluetooth devices. This includes security aspects and the control and negotiation of Baseband packet sizes. Furthermore, it controls the power modes and duty cycles of the Bluetooth radio device, and the connection states of a Bluetooth unit in a piconet. Logical Link Control and Adaptation Protocol (L2CAP). The protocol of L2CAP provides connection-oriented and connectionless data services to the upper layer protocols over the Baseband, with protocol multiplexing capability, segmentation and reassembly operation, and group abstractions, which permits higher level protocols and applications to transmit and receive L2CAP data packets. L2CAP is defined only for ACL links. Service Discovery Protocol (SDP). Using SDP to discover services is a crucial part of the Bluetooth framework and provides the basis for all the usage models. SDP query device information, services information, and the characteristics of the services, according to which a suitable connection between two or more Bluetooth devices can be established.

124 3

Security Framework

The Bluetooth technology provides security measures at both the application layer and the link layer. Besides there are two kinds of inherent features that make attacks more difficult. A hop selection mechanism of up to 1600 hops/sec is employed to avoid the interference from external or other piconets. An automatic output power adaptation scheme is also included in the standard for the low power consumption of light-weight mobile devices, which can reduce the radio spread range for data transmission exactly according to requirements based on the detected intensity. 3.1

Basic Definitions

A total of three different information security objectives are to be reached one or all. Confidentiality means that the data can only be used by authorized users and/or parties. Integrity means that the data cannot be modified during transfer and stored by adversaries. Availability means that the data is always available for authorized use. Typical attacks to a wireless network include DoS (Denial-of-Service), man-inthe-middle, spoofing, impersonating, session hijacking, eavesdropping, etc. Bluetooth launches three main techniques to achieve security features, including: • • •

3.2

Encryption. The process of transforming data into a form that it cannot be understood without a key. Both data and control information can be encrypted. Authentication. The process of verifying 'who' is at the other end of the link. Authentication is performed for both devices and users. Authorization. The process of deciding if a device is allowed to have access to a service. Authorization always includes authentication. Security Modes

Each Bluetooth device can operate on one of the 3 security modes. Mode 1 is a nonsecure mode in which a Bluetooth device shall never initiate any security procedure. Mode 2 is service-level enforced security where a device does not initiate security procedures before channel establishment at L2CAP level, and whether to initiate or not depends on the security requirements of the requested channel or service. Mode 3 is a link-level enforced security in which a Bluetooth device shall initiate security procedures before the link set-up at the LMP level is completed. Accordingly, two levels of Bluetooth security scheme can be identified, as follows: •

•

Link-level security, corresponding to security mode 3. The Bluetooth device initiates security procedures before the channel is established. This is the builtin security mechanism and it is not aware of service/application layer security. Service-level security, corresponding to security mode 2. The Bluetooth device initiates security procedures after the channel is established, i.e. at the higher layers. This is a kind of add-in mechanism and is regarded as a practical issue.

125

3.3

Security Levels

Bluetooth allows different security levels to be defined for devices and services. Two security levels can be defined for a device. A trusted device has unrestricted access to all or some specific services. Basically this means that the device has been previously authenticated and marked as "trusted". An untrusted device has restricted access to services. Usually the device has been previously authenticated but has not been marked as "trusted". An unknown device is also an untrusted device. Three levels of service security are allowed to be defined so that the requirements for authorization, authentication, and encryption can be set independently, including services that require authorization and authentication, services that require authentication only, and services open to all devices. These three security levels can be described by using the following three attributes: • • •

4

Authorization Required: access is only granted after an authorization procedure. Only trusted devices would get automatic access. Authentication Required: the remote device must be authenticated before being able to connect to the application. Encryption Required: the link between the two devices must be encrypted before the application can be accessed. Link-level Security

Figure 2 illustrates the link-level security framework of Bluetooth. In the figure, one of the two Bluetooth devices (the claimant) tries to reach the other one (the verifier). Generally four parts exist in the whole scheme as shown top down in the figure.

First Time In Operation (icncratc Iniuali/alion Key

First Time Handshake

I

Aulhenliculiun (Initialization Key)

1 Generate/Get Link Key

Following Handshakes

Data

:nt Link Key)

J

Transmission"! Figure 2. Bluetooth link-level security scheme.

126

4.1

Key Management Scheme

Key management scheme is used to generate, store, and distribute keys, which compasses the first step of each of the four parts in Figure 2. Basically, Bluetooth security scheme is based on symmetric key cryptography, i.e. a private key called link key is shared between two or more parties. The link keys can have different lifetimes. A semi-permanent key can be used after the current session is terminated, while a temporary key is valid only until the current session is over. A total of four types of link keys have been defined, as shown in Figure 3. The initialization key is used only during the initialization process. The unit key is generated once at the installation of the unit. The combination key is derived by both units for services that require more security. The master key, generated by the master device, is used when the master wants to broadcast messages. There is also a Bluetooth PIN used for authentication and to generate the initialization key before exchanging link keys.

X

llllllllllllllll

\ Initialization Kev

Figure 3. Bluetooth key structure.

4.2

Authentication Scheme

The Bluetooth authentication scheme uses a challenge-response strategy in which a 2-move protocol is employed to check a claimant's knowledge of a secret key, as shown in Figure 4. Claimant (Unit B)

Verifier (Unit A) AU_RANEk

Ex

BDjVDDRi Utile k&y

SRES

SEES

Figure 4. Challenge-response for the Bluetooth authentication.

* Ate

127

If authentication fails, a certain waiting interval must pass before a new attempt can be made. The waiting interval will increase exponentially. This is to prevent an intruder to repeat the authentication procedure with different keys. 4.3

Encryption Scheme

Figure 5 shows the encryption procedure. The encryption key (Kc) is generated from the current link key. There are several encryption modes in which broadcast messages and individually addressed traffic can be either encrypted or not, depending on whether a device uses a semi-permanent link key or a master key. Unit A (master)

Unit B

HP \|)|)R

Ml)

EN_RANDA ( I . -. 1

^

-•

-,

\1J1JK-

—

( '.- k.

k

»•

Data D.

1

* ff}4 -

Figure 5. Encryption procedure.

5

Service-level Security

This section describes the practical issues involved in the implementation of security mechanisms, i.e. an approach for a flexible security architecture built on top of the link-level security features of Bluetooth. More information can be found in [6]. Figure 6 illustrates the general security architecture. The key component in the architecture is a security manager, with the following functions: • • • •

Store security-related information on both services and devices into corresponding service and device databases. Grant or refuse access requested by protocol implementations or applications. Command the link manager to enforce authentication and/or encryption before connecting to the application, using the HCI. Query PIN entry to set-up trusted device relationship.

Employing such a centralized security manager is flexible to implement different access policies and easy to add new policies without affecting other parts.

128

Moreo¥er, the security manager acts as a bridge to join application level and linklevel security controls together and thus helps in providing end-to-end security. I

User Interface

I

Figure I. Bluetooth security architecture.

Authentication should be performed after determining what the security level of the requested service is. That is to say, the authentication can only be performed when a connection request to a service (SCO link) is submitted. 6

Vulnerability Discussion

Although the Bluetooth network system is relatively secure, by employing the schemes described above, there are still a number of weaknesses in the standard. Several pieces of research on addressing security flaws present in Bluetooth networks have been carried out [7-10]. Table 1 lists the detailed description and analysis of Bluetooth vulnerabilities, together with possible attacks/risks and countemieasures corresponding to each of the weakness issues. The table concerns the evaluations on all of the Bluetooth security specific protocols, including general security scheme, key management, authentication, encryption, and authorization. It is worthy to note that the entire framework of the Bluetooth security is acceptable. The weaknesses of the general Bluetooth protocols come from the wireless nature, ad hoc nature, and device address scheme. For security specific protocols, weaknesses fail into methods for PIN code, random number generation, unchangeable unit key, and security manager. More security can be expected through the employment of high-level security schemes by protocols anchor applications upon the Bluetooth.

129 Table 1. Vulnerability analysis of the Bluetooth standard.

Weakness The quality of pseudorandom number generator is undetermined.

Attack / Risk Guess the generator implementation or the generated pseudorandom number.

Countermeasure Statistical tests to detect non-repeating and randomly generated requirements.

PIN key is too short and default PIN is all zero.

Easy to exhaustively search or guess PIN key.

Increase the PIN code length.

Need to physically enter PIN code to devices.

Inconvenient PIN code input.

Application level key agreement software.

Initialization key is too weak.

Depend only on RAND and PIN which are both unsafe.

Employ new strong initialization key generation scheme.

Unit key is reusable, and is public to the other side once used.

Calculate encryption key or impersonate other devices with their unit key.

Use unit key as input to generate a random key. Use a key set instead of only one unit key.

Shared master key.

Impersonating or disclosing. Device embezzling.

Change broadcast scheme. Application level security and employ user authentication.

Disabling authentication attempts from legitimate devices. Shortcut attack: guess the contents of E0.

Encrypt device address. Limit the entry number of the list. Replace the cipher with other advanced scheme.

Encryption abort. Use too short a key. Security manager stands idle, no security for legacy applications.

Globe agreement on minimal key length. Add a Bluetooth-aware "adapter" application for the legacy application.

No separately defined authorization for services.

No service-related flexible device trusting assignment.

Modify the security manager and the registration processes.

Unidirectional access check but bi-directional traffic.

Malicious verifier attacks claimant by nasty messages.

Access check at all the phases and mutually. Check-consistent data flow direction.

No user authentication.

Repeating attempts for authentication. Weak E0 stream cipher. Negotiable key length. Leak support for legacy applications.

130 7

Conclusions

Although more and more qualified products emerge, at present Bluetooth is still more a laboratory technology to be studied than a widely used supporting technique for multitudinous products. The protocol is still in its research phase partly because of the security problems. Since the Bluetooth security scheme is reasonably robust to applications with less security requirements, the final features may depend more on the implementation than significant changes to the specification. Based on the original design goal of cable replacement, Bluetooth is more suitable to short-range and small-size wireless personal area networks than for connecting with outside public networks, comparing to e.g. WLAN. To applications such as large ad hoc networks and outside interconnection access, high level security schemes should preferably be enforced for complementation. Examples include e.g. IPSEC for IP, secure routing protocols, distributed secret schemes, etc.

8

Acknowledgments

The financial support from the National Technology Agency of Finland (Tekes) is gratefully acknowledged. References 1. Bluetooth SIG, Specification of the Bluetooth System: Volume 1, Core, Version 1.1, Feb. 22, 2001. 2. Bluetooth SIG, Specification of the Bluetooth System: Volume 2, Profile, Version 1.1, Feb. 22,2001. 3. IEEE 802.15 Working Group for WPANs, http://ieee802.org/15/. 4. Mettala R., Bluetooth Protocol Architecture: Version 1.0, Bluetooth White Paper, Document # l.C. 120/1.0, Aug 25,1999. 5. Haartsen J. C , The Bluetooth Radio System, IEEE Personal Communications, Feb. 2000, vol. 7, no. 1, pp. 28-36. 6. Muller T., Bluetooth Security Architecture: Version 1.0., Bluetooth White Paper, Document # l.C.l 16/1.0, July 15,1999. 7. Dasgupta K., Bluetooth Protocol and Security Architecture Review, online report, http://www.cs.utk.edu/-dasgupta/bluetooth/ 8. Vainio J. T., Bluetooth Security, Online report, May 25, 2000, http://www.niksula.cs.hut.fi/~jiitv/bluesec.html. 9. Jakobsson M. and Wetzel S., Security Weaknesses in Bluetooth, online report, http://www.bell-labs.com/user/markusj/bluetooth.pdf 10. Candolin C , Security Issues for Wearable Computing and Bluetooth Technology, Online report, http://www.cs.hut.fi/Opinnot/Tik86.174/btwearable.pdf

YOUR 802.11 WIRELESS NETWORK HAS NO CLOTHES WILLIAM A. ARBAUGH, NARENDAR SHANKAR AND Y.C JUSTIN WAN University Of Maryland College Park, MD 20742 USA E-mail: {waa, narendar, [email protected] The explosive growth in wireless networks over the last few years resembles the rapid growth of the Internet within the last decade. To protect the internal resources, organizations usually purchased and installed an Internet firewall. We believe that the current wireless access points present a larger security problem than the early Internet connections. A large number of organizations, based on vendor literature, believe that the security provided by their deployed wireless access points is sufficient to prevent unauthorized access and use. Unfortunately, nothing could be further from the truth. While the current access points provide several security mechanisms, our work combined with the work of others show that ALL of these mechanisms are completely ineffective. As a result, organizations with deployed wireless networks are vulnerable to unauthorized use of, and access to, their internal infrastructure. From the time when we had found out these problems to now, there have been proposals in the IETF for a new standard with better security mechanisms for wireless communication. These had been based on our findings and on another research group findings. However this involves changes to access points, which effectively involves changes to the current infrastructure. We believe that for the current installed base an interim solution, which works with the current deployed base is mandatory. This paper primarily talks about the some of the security flaws in wireless authentication that we had found and introduces the idea of a simple higher layer solution for most wireless security problems.

1

Introduction

Organizations are rapidly deploying wireless infrastructures based on the IEEE 802.11 standard [1]. Unfortunately, the 802.11 standard provides only limited support for confidentiality through the wired equivalent privacy (WEP) protocol, which contains significant flaws in the design [2,3]. Furthermore, the standards committee for 802.11 left many of the difficult security issues such as key management and a robust authentication mechanism as open problems. As a result, many of the organizations deploying wireless networks use either a permanent fixed cryptographic variable, or key, or no encryption what so ever. This fact, coupled with the fact that wireless networks provide a network access point for an adversary (potentially beyond the physical security controls of the organization), creates a significant long-term security problem. Compounding this is the fact that the access control mechanisms available with current access points contain serious flaws such that an adversary can easily subvert them.

131

132

Organizations over the last few years have expended a considerable effort to protect their internal infrastructure from external compromise. As a result, the organizations have canalized their external network traffic through distinct openings protected by firewalls. The idea is simple. By limiting external connections to a few well protected openings, the organization can better protect itself. Unfortunately, the deployment of a wireless network opens a ~back door" into the internal network that permits an attacker access beyond the physical security perimeter of the organization. As a result, the attacker can implement the Sparking lot" attack, where the attacker sits in the organization's parking lot and accesses hosts on the internal network. Ironically in some cases, the existence of the firewall may make the organization's hosts more vulnerable to the attacker because of the mistaken premise that the hosts are immune from attack and potential compromise. This paper describes the flaws in the two access control mechanisms that exist in access points built using Orinoco/Lucent 802.11 Wavelan PCMCIA cards, and a simple eavesdropping attack against the 802.11 specified shared key authentication mechanism. Exploiting these flaws when encryption is not enabled permits an adversary immediate access to the wireless network and most likely the organization's local area network as well. The use of encryption prevents an adversary from gaining immediate access, but combining our attacks with the weaknesses found in WEP by others provides such access [2,3]. The next section presents a short overview of the 802.11 wireless standard. This is followed an overview of the 802.11 security mechanisms, and Lucent's proprietary extension for access control. The next section describes attacks against the only two access control mechanisms available in most current access points, and an attack against the 802.11 standard shared key authentication mechanism. Finally, we conclude the paper with some insight into a higher layer authentication and key management scheme which can be used for wireless networks which exist currently. 2

802.11 Wireless Networks

802.11 wireless networks operate in one of two modes: ad-hoc or infrastructure mode. The IEEE standard defines the ad-hoc mode as Independent Basic Service Set (IBSS), and the infrastructure mode as Basic Service Set (BSS). In the remainder of this section, we explain the differences between the two modes and how they operate. In ad hoc mode, each client communicates directly with the other clients within the network. Ad-hoc mode is designed such that only the clients within transmission range (within the same cell) of each other can communicate. If a client in an ad hoc

133

network wishes to communicate outside of the cell, a member of the cell MUST operate as a gateway and perform routing. In infrastructure mode, each client sends all of it's communications to a central station, or access point (AP). The access point acts as an Ethernet bridge and forwards the communications onto the appropriate network- either the wired network, or the wireless network. Prior to communicating data, wireless clients and access points must establish a relationship, or an association. Only after an association is established can the two wireless stations exchange data. In infrastructure mode, the clients associate with an access point. The association process is a two -step process involving three states: 1. Unauthenticated and unassociated 2. Authenticated and unassociated and 3. Authenticated and associated. To transition between the states, the communicating parties exchange messages called management frames. We will now walk through a wireless client finding and associating with an access point. All access points transmit a beacon management frame at fixed interval. To associate with an access point and join a BSS, a client listens for beacon messages to identify the access points within range. The client then selects the BSS to join in a vendor independent manner. For instance on the Apple Macintosh, all of the network names (or service set identifiers (SSID)) which are usually contained in the beacon frame are presented to the user so that they may select the network to join. A client may also send a probe request management frame to find an access point affiliated with a desired SSID. After identifying an access point, the client and the access point perform a mutual authentication by exchanging several management frames as part of the process. The two standardized authentication mechanisms are described in later. After successful authentication, the client moves into the second state, authenticated and unassociated. Moving from the second state to the third and final state, authenticated and associated, involves the client sending an association request frame, and the access point responding with an association response frame. After following the process described in the previous paragraph, the client becomes a peer on the wireless network, and can transmit data frames on the network.

134 3

802.11 Standard Security Mechanisms

The 802.11 standard provides several mechanisms intended to provide a secure operating environment. At least one major vendor has implemented, authentication via the Extensible Authentication Protocol (EAP) [4]. The mechanism used is called Lightweight EAP (LEAP), which is a firmware based solution of EAP. We discuss later in the paper the problem of such a mechanism, which is the fact that it ties authentication with key management. In this section, we describe each of these mechanisms as well as a Lucent proprietary method. 3.1

Wired Equivalent Privacy protocol

The Wired Equivalent Privacy (WEP) protocol was designed to provide confidentiality for network traffic using the wireless protocol. The details of the algorithm used for WEP are beyond the scope of this paper. However, work by Walker and more recently by Borisov, Goldberg, and Wagner demonstrates that WEP, when used without a short key period, provides limited confidentiality [2,3], and possible misuse of the network. 3.2

Open System Authentication

Open system authentication is the default authentication protocol for 802.11. As the name implies, open system authentication authenticates anyone who requests authentication. Essentially, it provides a NULL authentication process. Experimentation has shown that stations do perform a mutual authentication using this method when joining a network, and our experiments show that the authentication management frames are sent in the clear even when WEP is enabled. 3.3

Shared Key Authentication

Shared key authentication uses a standard challenge and response along with a shared secret key to provide authentication. The station wishing to authenticate, the initiator, sends an authentication request management frame indicating that they wish to use "shared key" authentication. The recipient of the authentication request, the responder, responds by sending an authentication management frame containing 128 octets of challenge text to the initiator. The challenge text is generated by using the WEP pseudo-random number generator (PRNG) with the "shared secret" and a random initialization vector (IV). The IV is always sent in the clear as part of a WEP protected frame. Once the initiator receives the management frame from the responder, they copy the contents of die challenge text into a new management frame body. This new management frame body is then encrypted with WEP using the "shared secret" along with a new IV selected by the initiator. The encrypted

135

management frame is then sent to the responder. The responder decrypts the received frame and verifies that the 32-bit CRC integrity check value (ICV) is valid, and that the challenge text matches that sent in the first message. If they do, then authentication is successful. If the authentication is successful, then the initiator and the responder switch roles and repeat the process to ensure mutual authentication. The entire process is shown in figure 1, and the format of an authentication management frame is shown in figure 2. The format shown is used for all authentication messages. Initiator

Responder

Authentication Request Authentication Challenge Authentication Response Authentication Result

Figure 1. Mutual Station Authentication using shared keys 2 Frame Control

6 Duration

6

6

Dest. Addr

Source Addr

2 BSSID

2

0-2312

Seq Frame Body Num

4 (size octets) FCS

Mgmt Frame

Auth Frame Algo Num (size octets) Figure 2.

Seq Num

2

2

Status Code

Element Length ID

Frame Body 128

Authentication management frame

The value of the status code field is set to zero when successful, and to an error value if unsuccessful. The element identifier identifies that the challenge text is included. The length field identifies the length of the challenge text and is fixed at 128. The challenge text includes the random challenge string.

136 3.4

Closed Network Access Control

Lucent has defined a proprietary access control mechanism called Closed Network [5]. With this mechanism, a network manager can use either an open or a closed network. In an open network, anyone is permitted to join the network. In a closed network, only those clients with knowledge of the network name, or SSID, can join. In essence, the network name acts as a shared secret. Table 1 shows the possible values and when the challenge text is included based on the message sequence number.

Sequence

Status code

Challenge Text

WEP used

Reserved Status Reserved Status

Not present Present Present Not present

No No Yes No

Table 1. Message Format based on Sequence

3.5

Access Control Lists

Another mechanism used by vendors (but not defined in the standard) to provide security is the use of access control lists based on the ethernet MAC address of the client. Each access point can limit the clients of the network to those using a listed MAC address. If a client's MAC address is Usted, then they are permitted access to the network. If the address is not Usted, then access to the network is prevented. 3.6

Key Management

Key management is a misnomer with respect to 802.11 as it is left as an exercise for vendors. As a result, only a few of the major vendors have implemented any form of key management or key agreement in their high-end products. Unfortunately, none of the vendors provide sufficient information to determine the level of assurance provided by their product. Worse, in some cases, the details that are available indicate that the vendors vvsolution" worsens the problem by using protocols with well-known vulnerabilities, e.g. un-authenticated Diffie-Hellman key agreement. The 802.11 standard does, however, provide for two methods for using WEP keys. The first provides a window of four keys. A station or AP can decrypt packets enciphered with any one of the four keys. Transmission, however, is limited to one of the four manually entered keys- the default key. The second method is called a

137

key mappings table. In this method, each unique MAC address can have a separate key. The size of a key mappings table should be at least ten entries according to the 802.11 specification. The maximum size, however, is likely chip-set dependent. The use of a separate key for each user mitigates the cryptographic attacks found by others, but enforcing a reasonable key period remains a problem as the keys can only be changed manually. 4

Weaknesses in Current Access Control Mechanisms

This section describes the weaknesses in the access control mechanisms of currently deployed wireless network access points. 4.1

Lucent's proprietary access control mechanism

In practice, security mechanisms based on a shared secret are robust provided the secrets are well-protected in use and when distributed. Unfortunately, this is not the case with Lucent's proprietary access control mechanism. Several management messages contain the network name, or SSID, and these messages are broadcast in the clear by access points and clients. The actual message containing the SSID depends on the vendor of the access point. The end result, however, is that an attacker can easily sniff the network name- determining the shared secret and gaining access to the "protected" network. This flaw exists even with WEP enabled because the management messages are broadcast in the clear. 4.2

Ethernet MAC Address Access Control Lists

In theory, access control lists provide a reasonable level of security when a strong form of identity is used. Unfortunately, this is not the case with MAC addresses for two reasons. First, MAC addresses are easily sniffed by an attacker since they MUST appear in the clear even when WEP is enabled, and second most all of the wireless cards permit the changing of their MAC address via software. As a result, an attacker can easily determine the MAC addresses permitted access via eavesdropping, and then subsequently masquerade as a valid address by programming the desired address into the wireless card— by-passing the access control and gaining access to the "protected" network. 4.3

Shared Key Authentication Flaw

The current protocol for shared key authentication is easily exploited through a passive attack by the eavesdropping of one leg of a mutual authentication. The attack works because of the fixed structure of the protocol (the only difference

138

between different authentication messages is the random challenge), and the previously reported weaknesses in WEP [2,3]. The attacker first captures die second and third management messages from an authentication exchange, see table 1. The second message contains the random challenge in the clear, and the third message contains the challenge encrypted with the shared authentication key. Because the attacker now knows the random challenge P, the encrypted challenge ciphertext, C, and the public IV, the attacker can derive the pseudo-random stream produced using WEP, WEPK' ^PR, with the shared key, K, and the public initialization variable, IV, using equation 1. WEPK'IYPR = C © P (1) The size of the recovered pseudo-random stream will be the size of the authentication frame, (see figure 6) because all elements of the frame are known: algorithm number, sequence number, status code, element id, length, and the challenge text. Furthermore, all but the challenge text will remain the same for it ALL authentication responses. The attacker now has all of the elements to successfully authenticate to the target network— without knowing the shared secret K. The attacker requests authentication of the access point it wishes to associate/join. The access point responds with an authentication challenge in the clear. The attacker, then, takes the random challenge text, R, and the pseudo-random stream, WEPK' ^PR , and computes a valid authentication response frame body by XOR-ing the two values together. The attacker then computes a new integrity check value (ICV) as described in Borisov et. al. [3]. Now, the attacker responds with a valid authentication response message, and he associates widi the AP and joins the network. We identified this flaw independently. After explaining it to Jesse Walker, we learned that the 802.11 committee was aware of the problem [6] Utilizing the network when WEP is enabled, however, requires the attacker to implement the WEP attacks [3]. 5

New vendor solutions and new standards

New vendor solutions like AS-2000 and Airlock have been released recently. The advantage of these solutions is mat the link-layer key is renewed on a per-user, persession basis. However, the key does not change widiin one session. Even a user is online for weeks, the link-layer key does not change. In other words there is no timed key management protocol. Also key management is tied to authentication, because the keys are renewed only when die user authenticates. Besides, these solutions do not provide a proper means of authenticating a disconnected user. A disconnected user is an authorized user of the network but has left the network temporarily. As shown earlier in the paper it is mandatory to have a WEP enabled to

139 thwart some of the attacks described by us. If WEP is enabled , such a user cannot even authenticate himself because he might not be having the correct window of WEP keys. It is not clear as to how such a user can even authenticate himself with the network. Fortunately, the 802.11 standards body is currently working on significant improvements to the standard [7]. But one problem with this solution is that it necessitates changes to access points and also to clients, which means that a change in the infrastructure is required. 6

An insight into our solution

We now describe the solution, which we have developed and deployed at the University of Maryland. Our solution uses DHCP options as a transport mechanism for wireless key management and authentication. Besides our solution also takes care of issues like timed key management and the disconnected user problem. 6.1 1. 2. 3. 4. 6.2 1. 2.

3.

4.

Design Issues Provide a good key management system for wireless LAN's using die DHCP services of the wired LAN's Work with the existing infrastructure Limit complexity (complexity breeds design and implementation errors). Solve the timed key management and disconnected user problems Solution details Use of DHCP authentication [8] for higher layer authentication. Use of a wireless re-key DHCP option for key management, which when set along with the authentication option, can be used to transport the WEP key which is encrypted with a key generated from authentication. The solution solves the disconnected user problem by using a two-door approach at the link layer (using two keys, one a long term WEP key used for entry into die network (and subsequently authenticate) and another a short term WEP key, which is used for communication). The relatively long-lived WEP key is denoted by A. The key used for communication is denoted by K. The solution provides timed key management by tying key management and DHCP leases. When a user/STA renews its IP address the corresponding WEP key is also renewed (if die wireless re-key option is set).

140

6.3 Protocol The wireless network consists of stations (STAs) trying to connect to a wired network using the access point (AP). The DHCP server is in the wired part of the network. All link layer traffic is encrypted using K (the current link layer key). Wireless traffic from the STAs is encrypted using K. The AP decrypts the wireless traffic (because it too has K) and forwards die traffic to the wired network. The idea to mitigate current WEP flaws is to keep changing the current key K. At the same time valid STAs of the network, who leave the network must be able to obtain the current key (if it has changed) when they rejoin the network. This is accomplished by a double-door entry mechanism where the STA gains entry and authenticates itself (using higher layer DHCP authentication) into the network using A. The frequency of use of A is small when compared to K and is hence considered to be a key with a longer lifetime. Both the AP and the STAs have a window of four WEP keys. They can listen on any of the four keys but can transmit only on one key. The access point listens on both K and A. The STAs, which are a part of the network have K and A. The valid STAs who do not have K (they had left the network and are now rejoining) can gain entry and authenticate themselves using A. After the STAs have authenticated themselves they obtain the current link layer key K. Apart from rejoining the network, regular timed key management takes place for all STA's currently in the network. In other words the current key K keeps changing frequently. We use DHCP as a transport mechanism for getting the new link layer key Kn. The basic idea is as follows: 1. When the client/STA joins/rejoins the network, it is assigned an IP address and is given the current link layer key K. The IP address is leased for a particular time period. This time period is set by organizational policy. Apart from this the STA is also given the next link layer WEP key Kn. 2.

All the clients in the network who have the current key renew their IP address (NOTE: the address does not need to change) depending on the lease time and in the process also get the new link layer key Kn.

Message formats for the wireless re-key options are provided in [9]

141

7

Conclusions

The combination of our results with those of Walker and Borisov et. al. demonstrates serious flaws in ALL of the security mechanisms used by the vast majority of access points supporting the IEEE 802.11 wireless standard. The end result is that ALL of the deployed 802.11 wireless networks are at risk of compromise. An interim short term mitigation (not a complete solution) is a robust key management system for WEP and a higher layer authentication system, and the use of higher level security mechanisms, e.g. IPSec. These mechanisms are a good interim solution till when the new standards get deployed. 8

Acknowledgements

The authors would like to thank Jesse Walker of the Intel Corporation, Mark Seiden of Securify, and Angelos Keromytis of the University of Pennsylvania for providing valuable comments on a draft of this work. References 1. "LAN MAN Standards of the IEEE Computer Society. Wireless LANmedium access control (MAC) and physical layer(PHY) specification. IEEE Standard 802.11, 1997 Edition," 1997. 12 2. Walker J., "Unsafe at any key size: an analysis of the WEP encapsulation," Tech. Rep. 03628E, IEEE 802.11 committee, March 2000. http://grouper.ieee.Org/groups/802/ll/Documents/DocumentHolder/0 362.zi%p. 3. Borisov N., Goldberg I., and Wagner D., "Intercepting Mobile Communications: The Insecurity of 802.11." http://www.isaac.cs.berkelev. edu/isaac/wep-faq.html. 4. Blunk L. and Vollbrecht J., "PPP Extensible Authentication Protocol (EAP)," Tech. Rep. RFC2284, Internet Engineering Task Force (IETF), March 1998. 5. Lucent Orinoco, User's Guide for the ORiNOCO Manager's Suite November 2000. 6. Walker J.,"Overview of 802.11 security." http://grouper. ieee.org/groups/802/15/pub/2001/Mar01/01154r0P802-15_TG3%-Overviewof-802-ll-Security.ppt, March 2001. 7. IEEE 802.11 Working Group, http://grouper.ieee.org/groups/802/ 11/index.html. 8. Droms R. and Arbaugh W., Authentication for DHCP messages, RFC 3118 June 2001. 9. Shankar N., Arbaugh W. and Zhang K., "Message formats for DHCP wireless re-key option ". To appear soon as a Hewlett-Packard Tech report.

Part 6 Power Control and Performance Evaluation

VERTICAL OPTIMIZATION OF DATA TRANSMISSION FOR ENERGY AWARE MOBILE DEVICES K.F. DOMBROWSKI, M. METHFESSEL, P. LANGENDORFER, H. FRANKENFELDT, I. BABANSKAJA, I. MATTHAEI AND R. KRAEMER IHP, Im Technologiepark 25, D-15236 Frankfurt(Oder), Germany email: dombro @ ihp-microelectronics, com A major problem for TCP connections over wireless links is that errors introduced by the wireless channel interfere with the TCP protocol, leading to reduced data rates and power wastage. Based on accurate simulations for the TCP and the IEEE 802.11 MAC protocol layers, we discuss recipes to optimize the transmission by adjusting the MAC parameters such as the Long Retry Limit and packet fragmentation. We will show the influence of these parameters on power consumption. Hardware/software co-design can also contribute to a power efficient design of mobile hand-held devices.

1

Introduction

The envisioned integration of fourth generation wireless and mobile systems with today's existing fixed-network internet requires small, power-efficient, easy-to-use mobile terminals with high processing power. New types of services will be available universally to consumers and also for industrial applications ranging from high priority/high data-rate transmission such as real-time video to low priority/low data-rate transmission such as email. Depending on the application the requirements in terms of data-rate, processing power, power efficiency, error resilience and latency are certainly divers. From a system viewpoint two main areas need to be worked on in order to minimize cost and maximize up-time of the hand-held device for all possible applications: Firstly, it is not enough to optimize the performance of the separate protocol layers independently. Instead, additional performance enhancement must be achieved by coordinating the optimization across several or all of the layers. This general principal of "vertical optimization" should permit substantial additional power saving by using higher layer semantic information to control the lower layer state machines. Secondly, a partitioning of system functionality into hardware and software blocks within a single-chip solution, where the hardware parts (including the RF front-end in the case of software radio) act as accelerators (or co-processors) for specific data processing tasks. This paper is based on a project aiming at the development of a single-chip wireless broadband communication system in the 5 GHz band, compliant with the Hiperlan/2 and IEEE 802.11a standards. The single-chip system includes the complex digital baseband and DLC (data link control) functionality, the RF front-

145

146 end, and hardware "engines" for the high layer protocols (TCP/IP), the "vertical manager" and an application processor (JAVA engine). Based on accurate simulation models, we will show first results of the influence of vertical layer management between the transport layer (TCP) and the medium access control (MAC) layer on data transmission over an IEEE 802.11 wireless link. We will also discuss hardware/software partitioning strategies in the transport/network (TCP/IP) layer and the IEEE 802.11 MAC layer. 2

Optimization of TCP and wireless MAC interaction

A major problem for TCP connections over wireless links is that errors introduced by the wireless channel interfere with the TCP protocol, leading to reduced data rates, retransmission and power wastage. Figure 1 shows the type of system under consideration here. TCP runs as end-to-end protocol on the mobile device and the server. The IEEE 802.11 MAC protocol controls the transmission via the wireless link. TCP works well for the Internet connection between base station and server, but cannot differentiate between a lost packet in the internet, initiating congestion control, and a packet lost on the air link, which should be retransmitted without

Fig. 1: The system under consideration here. TCP runs end-to-end. An IEEE 802.11 MAC protocol controls the transmission via the wireless link.

initiating congestion control. The task is to optimize the performance in terms of power consumption while remaining within the standards for either protocol. This task can be achieved by i) a static assignment of parameters tuned to an optimum for average channel conditions or ii) a dynamic adjustment of system parameters to allow the adaptation to time-varying channel conditions. Both methods can be chosen in today's available IEEE 802.11b products by setting a fixed transmission rate and maximum packet length or by using auto-fallback to adjust for degrading channels. A third way to achieve optimal performance of TCP over wireless links is the vertical approach by letting the layers interact. The idea is that a global optimization exploits the available degrees of freedom more effectively than separate optimizations in subspaces. If for example the MAC layer doesn't know whether the packet to be transmitted belongs to streaming or non-streaming

147 0.08

1.0E-05

1.0E-03

Fig. 2: Overall data rate as a function of the wireless channel bit error rate (BER) for various values of the MAC long retry limit.

data, it must use the same sub-optimal treatment for both. This suggests the advantage of an interaction between the application and the transport and MAC layers. Previous work by different authors have shown an improvement of TCP performance over lossy links by the introduction of reliable link layer protocols (e.g. [1], [2],[3]). Our aim here is to study the interaction between TCP and IEEE 802.11 MAC layers without the introduction of new or modified link layer protocols. We want to use this information to develop schemes which improve the overall performance in terms of throughput and power consumption. By introducing vertical interaction as needed, we can optimize specifically for TCP, ignoring any disadvantages arising for other protocols such as UDP. Our simulations were done in the BONeS tool set from Cadence. The main effort was to implement realistic TCP and MAC layers; the IP and LLC layers were included in a rudimentary form. TCP was coded in C++ and was restricted to the established state, since connection establishment plays only a minor role for the overall performance. All relevant algorithms (congestion control, fast retransmit, retransmit timer adjustment, etc.) were included. The MAC layer includes the full CSMA/CA channel access algorithm, based on a model for the PHY layer which takes into account the correct timings and delays. For the wireless channel, the simulation used the parameters of the IEEE 802.11 standard for the FHSS case with a data rate of 2 Mb/s. We simulated the interaction between TCP and MAC in two ways: i) We modified the MAC layer retransmission limit (denoted here by LRL:

148

LongRetryLimit) and ii) we studied the influence of TCP vs. MAC layer fragmentation. The main results of these investigations are summarized below. 2. /

Modifying the MAC retransmission

When a packet is lost on the wireless link (either by a collision or by channel error) the MAC layer retransmits the packet up to a maximal number of attempts (denoted here by LRL: aLongRetryLimit). When the limit is reached, the packet is discarded 20 BER , 3.0E-4

15

10

-» 2.0E-4 o e e—o

o

B E B

IS

nil

20

1.0E-4 1.0E-5 1.0E-6 40

60 retry limit

—, 80

100

Fig. 3: Total number of bytes transmitted over the wireless link as a function of the MAC long retry limit for various channel bit error rates.

and the procedure is restarted with the next TCP packet in the output queue. However, it is the job of TCP to make sure that every packet arrives sooner or later. When the MAC layer discards a packet, the consequence is that TCP must retransmit it, albeit at a later time. In other words, every packet must be transmitted repeatedly until it finally arrives, and the only question is which part of the protocol stack does this. By varying LRL in the simulation, we can shift the burden of retransmitting lost packets between the MAC and TCP layers. We display the results in two ways for simulations with a fixed packet payload length of 1000 bytes. Figure 2 shows the data rate as function of the BER for various choices of LRL. This shows that the value of LRL makes a large difference in the transmission rate. The best performance is reached when LRL is made large. Figure 3 shows the total number of bytes sent into the wireless channel, which is a measure of the energy

149 IEEE802.il, 1 Mb/s PHY-PR 96 PHY-HD

DIFS 128 i

1 32

DATA 800

HEAD 568

BO 375

SIFS 28 i ACK 240

IEEE 802.11, 2 Mb/s PHY-PR 96 PHY-HD

DIFS 128 i

SIFS 28

I32

BO 375

HEAD 284

DATA 400

ACK 184

IEEE 802.1 l a , 6 Mb/s (scale x 5) PHY-PR 16 PHY-HD 4 1 i

DIFS 34 BO 67

HEAD 100

SIFS

\6

DATA 232

ACK 44

Fig. 4: Transmission times in \is at the PHY level for 100 bytes of data. Shown are the DIFS, Backoff (BO), Phy-Preamble and -Header, protocol headers, data-payload, SIFS, and Ack times.

consumption. Interestingly, this turns out to be essentially independent of LRL. Thus (for a given packet length, coding scheme, and BER) the required energy is a constant, while the required time becomes significantly larger if TCP takes care of retransmissions. The conclusion is that we always lose a certain amount of energy as we throw the packet against the faulty channel until it finally gets through, but that we only make things worse if this is done by TCP and not the MAC layer. The optimal MAC strategy for TCP data streams could thus be called "never give up" (NGU); there is no point in ever dropping a packet and proceeding to the next one at the MAC level, since it has to be retransmitted anyway. Some comments regarding this strategy are necessary: 1. Even for a large number of MAC-layer retransmits, the total delay is small compared to the minimal TCP retransmission timeout. 2. The strategy can only be used for TCP data. For other types of data, it might well degrade the performance. This will be handled by the vertical manager.

150

3.

2.2

Our simulations were done for static channel conditions. However, the basic idea of shielding TCP from packet loss by increasing the effort at the MAC level should also be valid for time-varying channels. TCP vs. MA C Layer Fragmentation

In the case of good channel conditions, the only effect of reducing the packet length is to increase the overhead needed to transmit a certain chunk of data. But if the channel becomes bad, large packets are more likely to be damaged, increasing the number of attempts needed to transmit the packet correctly. This slows down the transmission and increases the power consumption. In order to minimize these effects it makes sense to adapt the packet length to the channel conditions [2]. This

1000

800

600 length (byte)

400

200

0

Fig. 5: Effective data rate for a transmission of 1 Mbytes of user data across the system of Fig. 1 at a transmission rate of 2 Mb/s over the wireless link as a function of the packet length fragmented in the TCP or the MAC layer for various BERs.

involves a tradeoff between the improved probability that the packet is transmitted correctly and the increased overhead. One question of interest is whether a reasonable compromise can be found which gives adequate performance over a wide range of BERs, or whether a significant further gain can be obtained by choosing the length dynamically depending on the error rate. The packet length can be changed either by reducing the size of a TCP packet, or by introducing fragmentation at the MAC level. The most important difference is the amount of overhead added to the total data amount by each newly-created packet. For example, for the IEEE 802.11 standard at 2 Mbit/s, Fig. 4 shows that

151

1000

800

600 400 Packet Length (Byte)

200

Fig. 6:Total number of bytes transmitted over the channel as a function of packet length and BER.

transmission of the overhead takes 1127 us, while the user data requires 4000 and 400 us for payloads of 1000 and 100 bytes, respectively. The percentage used to transmit actual user data thus reduces from 78% to only 8%. The point is that a reasonable degree of fragmentation quickly leads to a situation in which the user data portion is small compared to the overhead. Figure 4 shows this similarly in the IEEE 802.11a case. In contrast, the overhead added by MAC-layer fragmentation is smaller, since it avoids the higher-level headers, the SIFS delay, and the backoff slots. The effect of TCP and MAC fragmentation on the data throughput at various error rates is compared in Fig. 5. Performance is always better for MAC fragmentation as expected. While the overhead is acceptable down to payloads of about 200 bytes for the MAC case, TCP fragmentation throughput decreases steadily with reduced packet size. Thus, fragmentation at the MAC layer is clearly preferable, leading to an improvement by a factor of two to three in some cases. Furthermore, the curves show that a fixed payload length of around 300 gives a reasonable performance over a wide range of channel conditions when MAC fragmentation is used, whereas a fixed choice is difficult to find for TCP-layer fragmentation. Changing the Long Retry Limit has no influence on the power consumption, as seen in Fig. 3, only on the time needed to finally transmit all data. Changing the packet length does influence the power needed for the transmission on certain amount of data. This is seen in Figure 6, which shows the total number of bytes

152

transmitted over the channel for various channel conditions as a function of the packet length. For good channel conditions it is better to send long MAC packets, to keep the protocol overhead small. As the channel degrades fragmentation helps to save power. It is favorable to fragment on the MAC layer instead of the TCP layer since the overhead is smaller. 3

Hardware/software co-design

To open a broad market for consumer products, cost and up-time of mobile devices are critical design parameters. One way to realize a low-cost system is to reduce system complexity and implement all functions on a single chip. The single-chip solution is also advantageous in terms of power dissipation and performance as compared to multi-chip solutions. Within the single-chip functionality has to be partitioned into software and hardware blocks in order to achieve an optimum overall performance in terms of power consumption and processing speed. The dynamic activation/de-activation of certain blocks during data transmission and reception allows us to introduce efficient power reduction mechanisms. The importance of hardware/software partitioning and its advantages are outlined by the following examples. A central part of the IEEE 802.11a baseband processing is the Viterbi decoder for forward error correction. Simulations have determined the Viterbi algorithm to consume between 98-99.9% of all calculation power in the receiver path. The remaining 0.1-2% is shared by all other functional blocks. Implementing the Viterbi as a standard DSP solution would consume approximately 8W of power, whereas our Viterbi decoder implemented in UMC's 0.25um CMOS consumes only 625mW in the worst case at a clock frequency of 80 MHz [4]. Comparable to the Viterbi decoder in the baseband processing are en/decryption and CRC calculation in the DLC. We are currently implementing this functionality on an FPGA board where we obtain speed enhancements of roughly a factor ten as compared to a pure software implementation on a 33MHz Arm7 processor. The above mentioned examples of hardware implementations are quite obvious, since they require large processing power. A vertical optimization in terms of hardware/software co-design should be more, than implementing the obvious parts of a system in hardware. Accelerators for specific tasks which occur on several layers of the protocol stack will also yield considerable power reduction. One example of such an accelerator could be a "string manipulation processor" operating on the transport (TCP/UDP), network (IPv4/IPv6/ICMP/IGMP...), and DLC layers.

153

Summary and Conclusions In this paper, we have shown the improvement of TCP data transmission over a wireless link in terms of throughput and power efficiency by vertical protocol optimization. Realistic models of the TCP and IEEE 802.11 MAC protocols were used in order to obtain quantitative information. The results demonstrate that there is a major advantage if the MAC layer can be made aware of the type of data it is handling, i.e. data from a connection-oriented service (TCP) as opposed to an UDPtype service. The optimal behavior for a reliable stream is then a "never give up" strategy on the MAC layer. Even after a large number of unsuccessful attempts it makes no sense to discard the packet, since the reliable transport protocol has to retransmit the packet anyway, albeit with negative side effects such as congestion window closure or blockage until a timer expires. Results were also obtained concerning optimization of the performance of power consumption by fragmentation. Since the overhead of the protocols becomes a significant portion of the total data as packets get smaller, fragmentation on the MAC layer is preferable to that on the TCP layer. Thus, the best implementation of the "never give up" strategy would include a stepwise lowering of the data rate and fragmentation on the MAC level until the packet has been transmitted correctly. Any approach along this line leads to additional complications, when several streams of different types (reliable and unreliable) have to be supported at the same time. However, the results show that the gains from such vertical protocol interactions could make the effort worthwhile. References

1. Balakrishna H., Seshan S. and Katz R. H., Improving Reliable Transport and Handoff Performance in Cellular Wireless Networks, Wireless Networks 1 (1995), pp. 469-481 2. Lettieri P. and Srivastava M. B., Adaptive Frame Length Control for Wireless Link Throughput, Range, and Energy Efficiency, Proceedings of the IEEE INFOCOM (New York, 1998), pp. 564-571 3. Parsa C. and Garcia-Luna-Aceves J. J.: Improving TCP Performance over Wireless Networks at the Link Layer, Mobile Networks and Applications 5 (2000), pp. 57-71 4. Grass E., Tittelbach-Helmrich K., Jagdhold U., Troya A., Lippert G., Kriiger O., Lehmann J., Maharatna K., Dombrowski K. F., Fiebig N., Kraemer R. and Mahonen P., On the Single-Chip Implementation of a Hiperlan/2 and IEEE 802.11a Capable Modem, IEEE Personal Communications (in press)

RADIOSCAPE - SYSTEM DESIGN TOOL FOR INDOOR WIRELESS COMMUNICATIONS VIA THE INTERNET YOSHINORIWATANABE, HIROSHIFURUKAWA, KAZUHIRO OKANOUE AND SHUNTARO YAMAZAH Networking Research Labs., NEC Corp. 1-1, Miyazaki 4-chome, Miyamae-ku, Kawasaki, JAPAN 216-8555 Tel: +81 44 856 2255, Fax: +81 44 856 8494 E-mail: {watanayn \furuhiro\okanoue \yamazaki} @ccm.cl. nee. co.jp Without appropriate supports, trends for personal usages of wireless systems can be discouraged due to the nature of invisible communication areas. In 2.4GHz band, it is essential to take care of interference among several wireless devices. This paper proposes RADIOSCAPE, a new client-server type Internet service to support users' wireless network design by themselves. Key components of RADIOSCAPE are a server with an accurate and rapid radio propagation prediction engine and a user client with user-friendly GUI to design wireless networks. We introduce ray-launching method on a parallel processing technology into the server to achieve an accurate and rapid radio propagation prediction. Through comparisons between field measurements and prediction results, we confirm that the server can provide the predictions at about 5dB of the average received power error within 100 seconds for ordinary two-story houses. We have started the RADIOSCAPE ASP service in Japan.

1

Introduction

Several merits produced by wireless LAN (Local Area Network) systems, such as no new wire, easy mobility and so on, attract people to introduce the systems into their home and/or office areas. Recent inexpensive IEEE 802.11b[l] compliant products, that provide transmission rate up to 11Mbps in ISM (Industry, Science and Medical) band within 2.4GHz, have widely encouraged people to enjoy the merits. Moreover, in order to support PAN (Personal Area Network) services, such as wireless connections between a cellular phone and a headset, Bluetooth[2] compliant products, that work at the same frequency band as IEEE 802.1 lb, are also spreading into consumers. However, the wide spread of wireless communication devices into the general public has a possibility to cause serious outage area problems, where wireless communications are not available. The outage areas can be classified into the following three patterns based on their causes; • •

areas where received power is too low to provide expected communication qualities, areas where distortion by multipath fading is too severe to provide expected communication qualities even if received power is enough and

154

155 •

areas where interference among devices operated at the ISM band is too severe to provide expected communication qualities.

It can be easily supposed that the outage area problems cause users' inconveniences to design their wireless networks, e.g. to find out the most preferable wireless devices locations on trial and error basis. Especially in 2.4GHz band, IEEE 802.11b, Bluetooth devices and microwave ovens can operate simultaneously. The outage areas caused by interference among the devices will seriously surprise the users, especially people who do not have much knowledge of radio systems because they would have seldom encountered such situations so far. In order to help the users to build sophisticated radio environment, this paper proposes RADIOSCAPE, an assistant tool for wireless LAN setup in homes or offices. The proposed RADIOSCAPE supports users to design preferable locations of the devices by themselves easily through visual demonstrations of the outage areas. The superior feature originates in a novel architectural concept of RADIOSCAPE based on a client-server model via the Internet. We also believe that RADIOSCAPE is one of the most promising and practical approaches for the systems coexistence issue. First of all, this paper describes a concept of RADIOSCAPE system and its system architecture. Then we evaluate RADIOSCAPE through demonstrating estimated outage area results and required computation time performances for the prediction. Finally, we conclude the paper. 2

RADIOSCAPE System Concept

In order to make the outage areas clear in a target environment, it is the first step to grasp propagation characteristics over die target environment. There exist several tools capable of estimating radio propagation characteristics[3][4]. However, it seems to be difficult for ordinary users to use such tools, because the tools typically require a high performance stand-alone type workstation that is too expensive to be owned by the ordinary users. In order for ordinary users to be able to recognize the outage areas in their environments easily and inexpensively, RADIOSCAPE is designed based on a client-server model via the Internet. The RADIOSCAPE server has a rich database regarding radio systems such as transmission power, antenna patterns and so on, all of which are not familiar to users but essential to achieve correct estimation of radio condition. The server also has a high-performance propagation analysis engine that can predict the coverage area of homes or offices by estimating received power, multipath distortion and interference accurately and quickly in cooperation with the technical information. On the other hand, the RADIOSCAPE client comprises two software tools, namely a layout editor and a radio viewer. A user inputs his individual environments data through the layout editor by locating furniture, rooms, wireless devices including a microwave oven, wireless LAN access points and so

156

on. The radio viewer displays the radio condition in the target building analyzed by the server. The RADIOSCAPE client is designed on GUI (Graphical User Interface) basis so that ordinary users can operate the software easily without special knowledge regarding radio systems. Fig. 1 shows a wireless system design flow provided by RADIOSCAPE. The RADIOSCAPE server is equipped in the Internet, and the RADIOSCAPE client is distributed to each user and installed in his PC connecting to the Internet. First of all, a user initiates the layout editor included in the RADIOSCAPE client and inputs a target environment, such as room and furniture layouts, location and a type of each access point, e.g., 802.1 lb, Bluetooth. After the editing process, the user sends the input environment data to the RADIOSCAPE server. Receiving the data, the server begins to analyze radio conditions at respective observation points for the given environment and sends the analyzed results back to the client. When the user cannot accept the radio condition for the target environment, then he can change position of objects within the environment by re-locating the access points and/or furniture. The above process is iterated until the user obtains acceptable radio condition for the target environment. Wireless device vender

individual user

Ik

_S

wireless device characteristics - antenna pattern - packet lirror R
'LT

propagation anahsis - retxived pov-'1 ill iil'iii - multipath di'i«iin>.i - interference •••!•• . .*

RADIOSCAPE server

1> Web access

o

radio viewer

RADIOSCAPE client r User's decision

^

- viewable o u ' . . - corninimicali- 1

preferable No communication area? »* access point relocation and so on Yes

Fig. 1. Wireless communication system design flow in RADIOSCAPE system

Thus, RADIOSCAPE can hide all of the technical parameters unfamiliar to ordinary users in the server. Hence, the ordinary users can easily try RADIOSCAPE and enjoy high quality wireless communications. Moreover, it is much easier for RADIOSCAPE system to provide more accurate and quick propagation analysis by

157

improving the prediction algorithm in the server without any modifications to the clients. 3

RADIOSCAPE System Architecture

3.1

Basic Architecture

Fig. 2 shows the system architecture of RADIOSCAPE. Users connect the RADIOSCAPE server via HTTP using a WWW browser. The RADIOSCAPE server receives environment data files from a RADIOSCAPE client and sends corresponding result files to the client. Radio Propagation Simulator

Fig. 2. System Architecture of RADIOSCAPE

3.2

RADIOSCAPE Server

A main task of the RADIOSCAPE server is radio propagation prediction. In the prediction, the accuracy is naturally important. Moreover, it is also required for the server to perform the prediction in a short duration, since it is supposed that the server needs to handle a lot of environment data files from users simultaneously. 3.2.1

Approach for Accurate Radio Propagation Prediction

Indoor objects such as walls, floors, ceilings, furniture and so on, form various and complex radio propagation conditions. Prediction methods based on statistical propagation models cannot predicts indoor radio propagation conditions well, that has been applied for outdoor propagation predictions such as cellular systems. Hence, we apply ray-launching method[5,6] to the RADIOSCAPE server, that is based on geometrical optics approximation and requires less computation time than other deterministic prediction methods, for example, Finite Element method or Finite Difference Time Domain method[3,7].

158 The ray-launching method launches rays to the discrete directions around from a transmitter. When each ray intersects a scattering object, it is divided into two rays, reflected and transmitted rays. The rays are re-launched to the direction decided by geometrical optics from the intersection point. The division and relaunching are recursively repeated and the paths of the rays in the target environment are decided. The method traces these paths and predicts received power and delay spread at each observation point based on information of rays passing near the point. The tracing of the paths is continued until the power of a ray or the total number of reflections and transmissions reaches given thresholds. The accuracy can be improved by increasing the number of rays launched from the transmitter or tracing paths of rays with lower power. However, there is a tradeoff between the accuracy and computation time. 3.2.2

Approach for Rapid Radio Propagation Prediction

In ray-launching method, there exists no interaction between launched rays. Hence, a parallel processing mechanism can be applied to tracing the rays. We build the RADIOSCAPE server based on a PC clustering technique to shorten computation time. Moreover, the PC clustering technique can provide flexible computation power according to the number of PCs. The required computation power of the server can be decided by the number of RADIOSCAPE users. Hence, the flexibility of the computation power can realize an appropriate balance between hardware scale of the server and the number of RADIOSCAPE users. 3.3

RADIOSCAPE Client

Since the RADIOSCAPE service aims at the general public, intuitive operation is essential for the client software, and therefore a great attention is given to user interface of the client software. The RADIOSCAPE client comprises two software applications as already mentioned, namely, the layout editor and the radio viewer, each of which is in charge of composition of building layout and display of radio propagation conditions. 3.3.1

Layout Editor

Fig. 3(a) shows an appearance of the layout editor. The layout editor provides a rich object library that includes several kinds of objects that we can typically see in home and/or office environments, e.g. walls, floors, tables, shelves, and so forth. Using the editor, the users compose their own home/office environments by locating objects in adequate positions based on the layout. Radio propagation condition is influenced by size of objects and material of which the objects consist. We can arbitrarily set 3-dimensional size of objects. With respect to the material, a list of well-known materials is built in so that we just choose a right one to each object

159

from the list. In order to describe the detailed structure of the environments, it is possible to make doors, windows and holes within respective wall objects. Multiple access points are allowed to be located in a single building as shown in Fig. 3(a), where left- and right-hands small circles denote an IEEE802.11b access point and a Bluetooth transceiver, respectively. An environment data file produced by the editor is sent to the RADIOSCAPE server through a WWW browser and the server starts the analysis of radio propagation for the environment. 3.3.2

Radio Viewer

The radio viewer indicates radio propagation conditions at respective observation points in the input environment. There are two different representations of view, namely, detail mode and overview mode. The detail mode takes color representation so that we can comprehend the magnitudes of SINR (Signal to Interference + Noise power Ratio) and delay spread at each observation point by reading the colors. In the overview mode, the radio condition at each observation point is simply expressed by a color either of white, gray or black. Each of the colors denotes good, marginal and poor communication availabilities, respectively, and the color is determined taking into account both SINR and delay spread. In the overview mode, it is possible to evaluate the interference between multiple radio systems. Fig. 3(b) shows an example of the overview mode in case of the home environment shown in Fig. 3(a). The radio conditions shown in Fig. 3(b) are due to the IEEE802.11b access point at the left-hand side. The color shown in each observation point denotes the communication availability that a terminal in the point will encounter. Since the environment data for this case includes a Bluetooth master, the evaluation results in Fig. 3(b) accounts for interference influence from the Bluetooth master.

a) Lyaout editor Fig. 3. RADIOSCAPE client

b) Radio viewer

160 4 4.1

Evaluation of RADIOSCAPE System Characteristics Comparison of Measured and Predicted Received Power Levels

To evaluate received power levels prediction performances of the RADIOSCAPE server, we performed field measurements in a two-story wooden house with a floor area of 150 m 2 in Fig. 4 and compared the measured and predicted received power levels. As a transmitter, we used an IEEE 802.11b access point located at two positions indicated API and AP2 in the figure. The transmitted signals from API and AP2 are received at the numbered positions and we measured the received power levels. Fig. 5 provides a scatter plot of measurements and predictions regarding received power levels in this house. «and o are measurement results from API and AP2, respectively. From the graph, high correlation can be recognized between predictions and measurements. The error in the predictions is mainly due to some differences between the actual environment and the models described in the environment data files. The environment data files corresponding to Fig. 5 can include several errors; for example, all of the objects are approximately composed of only planes, small objects are omitted, and some objects can include a little error in location. In addition, some rays passing outside the house do not reenter in the house without allocating other buildings around the house, which often causes underestimations of received power in the house. Even though the error factors are included in the data files, the average error between predictions and measurements was 4.5dB. As a result of a series of trials using several buildings including home and office environments, we confirmed that the average error was about 5dB.

The first floor

The second floor

Fig. 4. Floor plans and observation points in the field measurements

161 -30 -40 -50

• API •

o

o A1*Z *>5

-60 -70

y/G

•

-90l= -90

o S So 1

-80

0

jr jS

-80

O

•

o

L

-70

. „.

-60

-50

-40

-30

Measured receipved power [dBm] Fig. 5. Comparison of measured and predicted received power

4.2

Comparison of Measured and Predicted Communication Availabilities

In order to evaluate predicted communication availabilities, we performed field measurements of ping loss rate. Fig. 6 shows the floor plan of an office used for the measurements. The office has an area of 600 m2 and includes a lot of partitions to be 1.3m or 1.5m in height. We set up an IEEE 802.11b access point and measured communication availabilities between a PC at several measurement points and the access points. Through the measurements, transmission rate and packet length are fixed to 11Mbps and 1460 byte, respectively. The measurement results are plotted with O and X in the figure. O and X denote ping loss rate less than 50% and more than or equal to 50%, respectively, between each point marked O and X and the access point. It can be acceptable for a user to be able to use his wireless device in areas where the prediction denotes as low communication availability. Hence, we tuned a communication availability decision algorithm to a fail-safe manner that can guarantee a user to enjoy high communication availability at least in areas where the predicted results denotes as high communication availability areas. As described in Fig. 6, no Xs are shown in white areas where the prediction indicates high communication availability areas. Thus, it can be seen that the predicted and measured results have a good matching to our fail-safe manner.

Fig. 6. Communication availabilities predicted by the RADIOSCAPE

162

4.3

Computation Time Performance

As described in 4.1, we assume that an object is approximately composed of only planes. According to our experiences, an ordinary two-story house consists of 1500 to 2000 planes. Fig. 7(a) shows computation time with respect to the total number of planes. As can be seen in Fig. 7(a), the computation time increases as the number of total planes increases. For an ordinary two-story house, however, it takes no more than 90 seconds for analysis. In this evaluation, we use a cluster server composed of four PCs with 1GHz Pentium HI processor. Fig. 7(b) shows computation time versus the number of PCs used in a PC clustering. According to this result, computation time ideally decreases in proportion to the number of PCs even though there would be actually some overhead required to the communication between PCs.

2000

4000

6000

8000

10000

12000

2

3

Number of Planes

Number of PCs

a) Computation time vs Number of planes

b) Normalized computation time vs Number of PCs

Fig. 7. Evaluation of computation time

Conclusion This paper proposed RADIOSCAPE, a web solution to assist ordinary users on setting up wireless communication systems such as wireless LANs. A key issue for RADIOSCAPE is a rapid and accurate radio propagation prediction. In order to overcome the issue, we combined a ray-launching method for high accuracy and a parallel processing architecture for rapidity. As for accuracy, we confirm that the average error of received power is 5dB through several comparisons between field measurements and predictions in 2.4GHz band. It is also shown that computation time for the prediction is less than 100 seconds for ordinary two-story houses and can be further increased by adding PCs to the parallel processing. The client-server based system model of RADIOSCAPE can provide users more accurate and rapid prediction services by enhancing only the server capabilities. This model can be fit to ASP (Application Service Provider) services very well and we have started an RADIOSCAPE ASP service for 2.4GHz band

163

systems in Japan. The concept of RADIOSCAPE is so general that we are prepared to apply RADIOSCAPE to other wireless systems such as IEEE 802.11a in 5GHz. Users will be able to easily enjoy the new features by downloading updated clients. References 1. IEEE 802.lib-1999 Supplement to Information technology — Telecommunication and information exchange between systems - Local and metropolitan area networks - Part 11: Wireless LAN Medium Control (MAC) and Physical Layer (PHY) Specifications: Higher Speed Physical Layer (PHY) Extension in the 2.4GHz band 2. Specification of the Bluetooth System Version 1.1, February 2001 3. Tetsuro IMAI, Yuichiro INUKAI, and Teruya FUJII, "Indoor Microcell Area Prediction System using Ray-tracing Method", IEICE Trans. Commun., vol. J83-B, no.ll, pp. 1565-1576, November 2000 4. http://www.ntt-at.co.jp/news/release/release63.html 5. J. P. Rossi, J. C. Bic, A. J. Levy, Y. Gabillet, and M. Rosen, "A ray launching method for radio-mobile propagation in urban area", in IEEE Antenn. Propagat. Sympos., London, Ontario, Canada, June 1991, pp.1540-1543 6. S. Y. Seidel and T. S. Rappaport, "Site-Specific Propagation Prediction for Wireless In-Building Personal Communication System Design", IEEE Trans. Veh. Technol., 43, 4, pp.879-891, 1994 7. Kazuhiro UEHARA, Tomohiro SEKI, and Kenichi KAGOSHIMA, "Indoor Propagation Calculation Considering Antenna Patterns Using Geometrical Optics Method", IEICE Trans. Commun., vol. J78-B-II, no.9, pp.593-601, September 1995

CAPTURE EFFECT IN IEEE 802.11 WIRELESS LANS ZORAN HADZI-VELKOV Faculty of Electrical Engineering, Karpos 2 bb, WOO Skopje, R. E-mail: [email protected]

Macedonia,

BORIS SPASENOVSKI Faculty of Electrical Engineering, Karpos 2 bb, 1000 Skopje, R. E-mail: boriss@cerera. etfukim. edu.mk

Macedonia,

The impact of frame capture over the IEEE 802.11 DCF is considered by two capture models based on pure Rayleigh-fading and combined Rician/Rayleigh-fading radio channels. For both models, we derived closed form solutions for the capture probability and the DCF saturation throughput. Simulations proved the Rician-faded channel to have similar impact over capture probability and network capacity as compared to the Rayleigh-faded capture model. The packet size threshold over which it is convenient to switch from Basic to RTS/CTS access scheme also proved sensitive to the capture effect. Generally, the sensitivity of an IEEE 802.11 network to capture depends on the receiver design.

1

Introduction

Today the wireless local access technology is essentially based on the IEEE 802.11 Wireless LAN standard [1], Although it is envisaged to provide high bandwidth to wireless stations in indoor environments, the radio channel introduces significant complexity to the design of wireless networks and analysis of their performances [2, 3]. The radio channel is exposed to multipath fading, which primarily results in increase of frame-error probability, but also introduces the capture effect. By properly modeling the multipath radio channel in an IEEE 802.11 ad-hoc network, the capture effect can be naturally modeled. In this paper, the frame capture effect is considered by introducing two capture models based on both the pure Rayleigh-fading and the combined Rician/Rayleighfading radio channel. Their impact over the capacity of IEEE 802.11 Distributed Coordination Function (DCF) is estimated analytically and verified by simulation. We also considered the packet size threshold over which it is convenient to switch from Basic to RTS/CTS access scheme with respect to the utilized capture models. The results have practical significance regarding the IEEE 802.11 access method implementation. This paper is organized as follows. The maximal capacity of the IEEE 802.11 DCF with regard to the two capture models is analyzed in section 2. Section 3 compares the analytical and simulation results of the maximum achievable throughput. Finally, the conclusions are provided in section 4. 164

165

2

Capture Model in Radio Channel

The capture effect is observed considering the basic direct-sequence spread spectrum (DSSS) physical-layer implementation of the IEEE 802.11 standard [1] in the 2.4 GHz ISM band. The frame transmission is performed at signaling rate of 1 Mbit/s using a fixed sequence of the 11-chip Barker code, transmitted at 11 Mchip/s, in order to spread die 1 Mbit/s PSK signal (i.e. spreading factor Sy = 11).

2.1

Saturation Throughput

A frame captures a receiver if frame's detected power Ps sufficiently exceeds the joint interfering power P„ of n interfering contenders. However, since there is a DSSS transmission involved, there is a performance improvement reflected in the processing gain through the implementation of the correlation receiver. The processing gain actually introduces a reduction of interference power by a certain factor g(5f), inversely proportional to the spreading factor S/. Assuming each chip in a frame is rectangularshaped, reduction factor can be expressed as in [4], i.e. g(S/) = 2/(3-5/). Thus, the conditional capture probability given » interfering frames (i.e. n interfering stations) is the probability of signal-to-interference ratio y= PJP„ exceeding value of z0 • g(Sj). This probability corresponds to the complementary distribution function Prob(y> z0-g(Sj)\ ri) = 1 - Fj(z0), where F/z 0 ) is the integral distribution function [5]. The capture effect is expected to produce throughput improvement, because of the finite possibility of successful transmission during simultaneous transmission of multiple colliding stations. As according to [2], the maximum achievable (saturation) throughput S ^ of the IEEE 802.11 DCF in ideal channel conditions can be expressed as: £

9 s

c

[f]

(!) p sue

where E[P] is the average packet payload size (in our case, E[P] = P, because all generated packet have fixed and maximal size of P — 2312 octets); Ptr is the probability that there is at least one transmission in the observed time slot; Psuc is the probability of a successful transmission assuming at least one station is transmitting; and a is a duration of an empty slot time. Ts is the average time the channel is sensed busy because of a successful transmission, and Tc is the average time the channel is sensed busy by each station during a collision. The values of Ts and Tc differ depending on the access scheme: Basic or RTS/CTS "handshake" transmission. Parameters P, Ts, Tc and crmust be expressed in same unit (bits or us).

166

Given N stations contending for a channel, the probability P fr can be expressed through the probability T of a station transmitting in a randomly chosen slot time. Actually, Plr is the probability of at least one transmission in the observed time slot assuming TVpotential transmitters: />,r=l-(l

(2)

•if.

The probability T depends only on the number of contending stations N, initial contention window Wand the maximal number of retransmissions (Retry-Limit) before a frame is discarded from its transmit queue. The Retty-Limit parameter is related to the random back-off mechanism (i.e. binary exponential back-off) and significantly influences the network throughput. Using the simulation model explained forthwith, we evaluated probabilities x for different values of N, given W = 8 and Retry-Limit = 5 (Table 1). Probability T can be used to define the probability of frame capture Pcapure as the following: (3)

:YJRiJ>rob(y>zog(Sf)\i)>

where Rt is the probability of /' interfering frames being generated in the observed time slot, namely R.=

( N \

ri+,(l-T)M-

(4)

7+1 Assuming at least one station is transmitting, probability of a successful transmission Psuc is the sum of the probability of exactly one station transmitting, and the frame capture probability ^capture- Thus, we have: P =

NT(l-ry

+P capture

sue

NT(l-r)N-'+JjRl-Prob(r>z0g(Sf)\i) i=i

,

1-0-T)" Table 1. Probability T estimated by simulation TV

W = 8 Retry-Limit = 5 T

10 20 30 40 50

0.076 0.051 0.041 0.035 0.031

(5)

167

2.2

Pure Rayleigh-Fading Channel with Capture

First, we used somewhat arbitrary assumption for no line-of-sight (LOS) among stations, so that the distribution of some signal's envelope at some receiver is assumed to be Rayleigh-faded. It means that the frame's instantaneous signal power Pt at the receiver is negative-exponentially distributed: (6) fp,(p) = — exp(--^-)> Po Po where p0 represents the local mean power, which is determined by path-loss effects during propagation. It means that the power of transmitted signal decreases with increasing distance to a certain path-loss exponent. The path-loss exponent for indoor channels in (picocells) typically ranges between 2 and 6.8. In order to estimate conditional capture probability Prob(y > z0-g(S/)| «), we assumed the power of each signal (of both the useful and each of the interfering frames) is Rayleigh-distributed as according to (6). Given equal local mean power p0 of all frames at the receiver, based on [6] we have

Prob{y>z0g{Sf)\n)

= \l[\ +

g{Sf)zJ.

(7)

Putting (3), (5) and (7) together, for the Psuc we obtained a closed form solution:

[\ + (\-T)g{Sf)-z0Y

-[\ +

g(Sf)-z0Y(\-xY

(8)

[l-a-Tn-ti+s^vz.r-1

Substituting (8) into (1), we obtained a closed form solution of the saturation throughput 5max of IEEE 802.11 DCF under influence of the capture effect in a Rayleigh-fading channel. —i——^

0.9 0.8

•".-*..

0.92 -

I

' • •

E

""'..

™ " - ^ -

•

—

r

^

'••'•r>

••> ^

£0.91

••

^ ^ ^

ations = 10

0.2

NoS ations = 30

* - -..

0.1 0

'

NoStalions = 30

^_ *^

1

9

(a) Basic

12

Z:W)

18

(b) RTS/CTS "handshake"

Figure 1. Theoretical saturation throughput vs. capture ratio in a Rayleigh-faded channel

168

Fig. 1 shows the curves of £max in function of capture ratio z0. Note that all contending frames at the receiver are assumed to have equal local mean power p0. Multiple curves correspond to Basic (Fig. la) and RTS/CTS "handshake" (Fig. lb) transmission scheme, respectively, and the number of stations appear as parameter in graphs. Probability T, appearing in (2) and (8), is chosen from Table 1. The values of other parameters appearing in (1), such as Tc and Ts, are identical to the ones of the simulation model explained forthwith. If Basic access scheme is utilized, it is obvious that presence of capture effect generates significant throughput improvement as z0 decreases below 24 dB. For example, given z0 = 6 dB and 10 stations within a single Basic Service Area (BSA), the peak throughput is estimated to around 90 % as opposed to 67 % given za—> °° (absence of capture). Conversely, the use of the RTS/CTS "handshake" access scheme contributes significantly to the robustness of the IEEE 802.11 network under capture. Similar to its influence over network insensitiveness to the number of stations and the hidden terminal effect, the RTS/CTS "handshake" mechanism allows only minor throughput increase of up to 3% as z0 decreases from 24 to 6 dB. 2.3

Combined Rican/Rayleigh-Fading Channel with Capture

Allowing for more realistic channel conditions, we assumed Rician-faded envelope statistics for the useful signal component with Rician factor K, which typically achieves values between 6.8 and 11 dB for indoor channels. This assumption suggests LOS presence between transmitter (of the successfully captured frame) and corresponding receiver. Given local mean power p0d at the receiver, the signal power P, of the useful frame is distributed as follows: , , v K +\ (K + l)Pl , 4K(K + \)p A (P) = exp[-K - ± *-] • /„ Pod

(9)

Pod

Pod

The interferers are still assumed to have Rayleigh-distributed envelope. Namely, n interfering signals have equal power levels pot, so that the pdf of the joint interference power P„ (obtained by incoherent addition of signals) can be expressed as: Pm (n-\)\

p,o;

Let us define F^z0) = Prob(P/P„ < z0-g(S/) |«) as the following: Fy(zo) = Prob(P„>

Z-—\„) = ]dx-fPt(x) Z

o •£(•*/)

o

]fK(y)dy

(1!)

W[v«(S/)]

because useful signal Ps and interfering signal P„ are statistically independent variables. Introducing (9) and (10) into (11), we have:

169

Fy(z0) =

1

0

\k

=j*-/,,(*).gl n-\

g(Sf)

1

k=0k\

-g(Sf)-z0

1

f ( 2 1 ) - ' expC-^L)^ />o; ("-1)!,/tz0Jg(S/,] Po, Pot

]dxfPi(x)

•/>,

• z0 • Poi

f exp v g(Sf)-zo-p0

-)*]**•/,,(*)• exp(-

gCS/Vz,, p0i

)dx

1 ^(-0) -^rJ/,,W-exP(-«^

(12)

where 0=1/ [g(5»Zo^0i]The integral behind the differential corresponds to the characteristic function of the signal power in a Rician-faded channel [7], which is: \fP (x) • exp(-
exp


(13)

Finally, for the conditional capture probability we obtain: 7* f

F(z1=T

•—

K+l K+l+Q-Pu

P
exp

K+l+(t>-Pod

(14) J) Hi*< s />*•>''»'l

Using some mathematical package, such as Mathematica, it is relatively easy to obtain closed form solution for the capture probability ^capture, introducing Prob{y> z0-g(SJ)\ri)= 1 - F^z0) into (3). It yields to the closed form solution for the probability of a successful transmission Psuc according to (4), and the saturation throughput Smax according to (1). Note that if K = 0 &nAp0i = Pod, (14) transforms into (7), so that the combined Rician/Rayleigh-fading capture model turns into the pure Rayleigh-fading capture model.

3

Performance Evaluation Model

We designed a C/C++ discrete-event simulator in order to verify the derived expressions for the capture probability Pcapture and saturation throughput Smax of the IEEE 802.11 DCF. We implemented the DCF within a single BSA, which consists of multiple randomly distributed wireless stations. We neglect the propagation delay, so that all the collisions can occur only due to simultaneous start of transmission of multiple stations. Depending on the access scheme, average times Tc

170

and Ts are chosen according to [2]1. Table 2 displays the relevant parameters of the simulation model according to the IEEE 802.11 standard. Table 2. Relevant system model parameters

Parameter

Default

Channel Rate PHY Preamble PHY Header MAC header ACK RTS CTS SIFS DIFS Slot Time rj RTS Threshold Retry Limit Transmit power

1 Mbit/s 144 symbols 48 symbols 34 octets 14 octets 20 octets 14 octets 20 us 50 us 20 us 200 octets 5 50 mW

The traffic generated from upper layers and delivered to the transmit queue in each station is Poisson-modeled. All stations are assumed to generate equal traffic load, so that the total generated traffic load can lead the system into saturation. In order to establish upper limit of the throughput, unless emphasized, all generated packets are assumed to have fixed and maximal possible length, i.e. P = 2312 octets. 3.1

Capture probability and network capacity: analysis vs. simulation

Simulations show that the impact of the capture over the throughput is negligible under conditions of low and moderate traffic loads. Once the network is lead into congestion, simulations confirm the analytical results (and Fig. 1) that the capture effect introduces significant throughput improvement as z0 decreases (below 24 dB) given Basic transmission is utilized, as compared to the peak network utilization in the absence of capture. The curves of Fig. 2 depict the capture probability PcaptuTe(N> 1) in function of capture ratio z0, assuming at least one interferer (N> 1), i.e. at least two contending stations. The curves are produced by simulation assuming Rician-distributed envelope of both useful and interfering signals, with realistic (and possibly mutually different) local mean powers at the receiver (depending on the propagation distance) and with identical Rician factors K (K= 0, 5, or 10). PHYpre/hdr + MAChdr + P + SIFS + ACK + DIFS PHYp^dr + MAChdr + P + DIFS RTS + SIFS + CTS + SIFS + PHYpre/hdr + MAChdl + P + SIFS + ACK + DIFS RTS + DIFS

171

i ; « • > ,

0.8

>* \\

=0 =5 =10

»0.6

[0.4

^S

0.2

12

z 1 0 5 (dB) 1 8

24

21

Figure 2. P^ptm^Ni 1) in a simulated Rician-faded channel with realistic (distance-dependent) local mean powers of all signals at the receiver

Additionally, comparing the two capture models introduced previously, Fig. 3 shows the Sm!a vs. z0, assuming: (1) pure Rayleigh-faded signals (K = 0), (2) combined Rician and Rayleigh-faded signals (useful signal is assumed to have K= 10), and (3) non-faded signals (absence of capture). The curves are produced presuming Basic access scheme, N= \0 stations and equal mean power level of both the useful signal and the interfering signals at the observed receiver (pod=pod1 0.9 I 0.8 » 0.7

"•-—C t

0 10 capture

^ "~ c ^ s 1

"

a. 0.6 .c 50.5 o £0.4 ^0.3 n £0.2

0.1 12

15

18

z 0 (dB)

21

24

Figure 3. Comparison of capture models with respect to the peak network capacity given equal local mean power of all signals at the receiver

It is obvious from Fig. 3 that the presence of the useful signal with LOS component (K - 10) generates peak capacity difference of only up to 4 % as compared to the pure Rayleigh-faded statistics of all signals. The simulation results match the analytical ones from Fig. 1 and Fig. 3, showing the close correspondence of curves relating to Rayleigh and Rician fading capture models under realistic

172 values of the local mean powers. Proximity of the three curves from Fig. 2 also sustain this conclusion. Given RTS/CTS "handshake" access is used, simulation curves of the peak throughput versus z0 almost completely match the ones from Fig. lb regardless of the Rician factor K. Identical conclusions apply for network with different number of stations. 3.2

Packet size threshold: RTS/CTS or Basic Access Scheme

In the view of the obviously significant impact of the capture effect over the network capacity, it would be interesting to realize its influence over packet size threshold over which Basic access mechanism should be switched to RTS/CTS "handshake" access scheme. As we established previously, RTS/CTS mechanism greatly contributes to the robustness of system performance under capture. However, it also degrades peak network capacity given shorter frames are transmitted. Under assumption of fixed packet payload size P, it is possible to quantify a packet size threshold by placing a condition where the saturation throughput under RTS/CTS access exceeded the saturation throughput under Basic access. This condition yields to the following expression [2]: n ^ /rprtslcts

/ n-i basic

iybasic\

rprtsl cts \

(15)

1-P„ 1200 NoStations = 10 NoStations = 30 NoStations = 50

r 400

I

200

\ -\\ >

-> •

12

"

.

—

-

. fdB) "

Figure 4. Capture effect also impacts the selection of the access method, Basic or RTS/CTS "handshake"

7; = 87 octets, and r c • r rts/cts = 1 4 Q c t e t s _ According to Table 2, Ts A closed form solution is attained introducing (8) into (15), which provides a certain insight on the dependence of packet size threshold with respect to the capture effect in a pure Rayleigh-fading channel. Thus, the curves of packet size threshold vs. capture ratio are depicted in Fig. 4. Thay have a practical significance in defining optimal operational parameters of an IEEE 802.11 network. basic

basic

173 4

Conclusions

In this paper, the influence of capture effect over the IEEE 802.11 DCF is regarded through the capture probability P'caplure considering two capture models based on two indoor channel fading conditions: pure Rayleigh-fading and combined Rician/Rayleigh-fading. For both models, we derived closed form solutions for the conditional capture probabilities, which combined with results from [2], yield to analytical solutions for the saturation throughput 5max in function of the capture ratio z0. The throughput improvement under Basic access scheme can vary depending on capture ratio, i.e. the receiver design. Simulations showed pure Rician-faded channel to have similar impact over capture probability and network capacity as compared to the Rayleigh-faded capture model. The packet size threshold over which it is convenient to switch from Basic to RTS/CTS access scheme also proved sensitive to the capture effect. Finally, it is important to emphasize that all the results refer to the capacity of the IEEE 802.11 DCF where the channel fading is involved only from the aspect of the capture effect. In order to evaluate the system capacity under real operating conditions, one must take the frame-error probability (depending on the modulation scheme, fading, ISI, noise and receiver structure), hidden terminal effect, etc. into consideration. References 1. IEEE Standard for Wireless LAN Medium Access Control and Physical Layer Specifications, IEEE 802.11b (Nov. 1999) 2. Bianchi J., Performance Analysis of IEEE 802.11 Distributed Coordination Function, IEEE J. Select. Areas Commun., vol. 18, No. 3 (March 2000) pp. 535-547 3. Hadzi-Velkov Z. and Spasenovski B., Performance Comparison of IEEE 802.11 and ETSI Hiperlan type 1 under Influence of Burst-Noise Channel, Proc. IEEE WCNC2000, Chicago, IL, USA (September 23-28, 2000) 4. Perez-Romero J., Alonso L.G., and Agusti R., Average Block Error Probability in reverse Link of a Packet DS/CDMA System under Rayleigh Fading Channel Conditions, IEEE Comm. Letters, vol. 4, No. 4 (April 2000) pp. 116-118 5. Lau C. T. and Leung C , Capture Models for Mobile Packet Radio Networks, IEEE Trans, on Commun., vol. COM-40 (May 1992) pp. 917-925 6. Dardari D., Tralli V. and Verdone R., On the Capacity of Slotted Aloha with Rayleigh Fading: The Role Played by the Number of Interferes, IEEE Comm. Letters, vol. 4, No. 5 (May 2000) pp. 155-157 7. Wand L. C. and Lea C. T., Co-Channel Interference Analysis of Shadowed Rician Channels, IEEE Comm.Letters, vol. 2, No. 3 (March 1998) pp. 67 - 69

Part 7 Medium Access Control

OPTIMISING THE POLLING SEQUENCE IN EMBEDDED ROUND ROBIN WLANS LACHLAN L.H. ANDREW AND RAVINDRA S. RANASINGHE ARC Special Research Centre for Ultra-Broadband Information Networks (CUBIN), Department of Electrical and Electronic Engineering The University of Melbourne, VIC 3010, Australia

{1.andrew,r.ranasirighe}@ee.mu.oz.au Tel. +61 3 8344 3816

Fax. +61 3 8344 6678

Transmitting variable-bit-rate real-time data on the uplink of a polled wireless local area network requires careful scheduling to achieve satisfactory performance and capacity. This is a "blind" scheduling problem, since the number and arrival times of packets at each remote station are not known by the scheduler. Embedded round robin (ERR) was recently proposed to address the problem of the high cost of polling idle stations. This paper presents an enhancement of ERR, least-recently-used-ERR, which outperforms the original ERR, especially when network load is high.

1

Introduction

In wireless local area networks, bandwidth management (i.e., packet scheduling) is crucial to providing acceptable quality-of-service (QoS) for real-time sessions having diverse traffic characteristics1. Real-time multimedia services are expected to form a major component of future network traffic. The IEEE 802.11 wireless local access network (WLAN) standard2 specifies that real-time data be transmitted by polling, but does not specify the order in which stations are to be polled. Scheduling downlink packet streams is essentially a centralized task, and is analogous to scheduling in wired networks. Many centralised scheduling algorithms have been proposed to achieve different QoS characteristics such as low delay3,4, low delay jitter5, fairness 6,7 ' 8 ' 9,10 , and maximum end user satisfaction11. All of these schemes require the scheduler to know details pertaining to each active session (i.e., packet arrival times, queue lengths, packet lengths etc.). These algorithms can be easily implemented in a centralised queueing environment since the required information is available. The uplink of wireless networks presents particular difficulties for packet scheduling, especially in the presence of non-uniform load. When a distributed queueing system is controlled using a centralised scheduler, the obvious mechanism of granting channel access to mobile stations is polling. One possible polling strategy is round robin12 (RR), which polls stations cyclically, regardless of the state of their queues. When RR is used to schedule service from a centralized queue, the scheduler can efficiently bypass stations that have no data to transmit. In contrast, the scheduler of a wireless network must poll a station in order to determine that it has no data to transmit. This overhead can be significant; for example, in IEEE 802.11 wireless LANs such a null poll takes about 5% of the time of a maximum length packet transmission2. If the scheduler is able to poll queues non-uniformly depending on their current load then it can save bandwidth. Most existing non-uniform polling schemes require explicit messaging. (See for example Reference 13.) This not only introduces transmission overhead but is also incompatible with many WLAN standards2. Other MAC protocols 177

178 have been proposed 14 ' 15 ' 16 to extend to wireless networks some of the fair queueing and deadline-ordered scheduling schemes popular in wired networks. These again generally achieve QoS at the expense of explicit messaging. Several schemes have been proposed to avoid null polls, notably embedded round robin17 (ERR). This algorithm is the basis for the wireless dual queue, which implements the dual queue algorithm11 for transient congestion avoidance. This paper will explore the potential of the embedded round robin paradigm to provide high quality of service under a range of conditions. After reviewing the original ERR algorithm, the least-recently-used ERR (LRU-ERR) algorithm will be described and its performance will be investigated under a range of conditions. The LRU-ERR algorithm aims to increase the proportion of packets being served without undue delay in the presence of high loads. 2

Embedded Round Robin

One of the major problems associated with scheduling distributed uplink queues is the lack of information available to the scheduler. Therefore the scheduler may poll potentially idle stations, wasting bandwidth. By knowing whether or not a station had further packets to transmit the previous time that it was polled, the scheduler can identify stations which are are guaranteed not to yield null polls. This concept is the basis for the embedded round robin (ERR) scheme17. Most polling protocols provide a single bit of feedback from remote stations indicating the presence of more data to transmit. For example, it is supplied by the the more d a t a bit of the IEEE 802.11 header2. Based on this feedback, traditional exhaustive round robin repeatedly polls a single station until it has no more data, before advancing to the next station in round robin order. This only risks polling an idle station when there are no stations known to have data, and hence minimises the number of null polls. A polling strategy with no null polls (and no packet discards) is "work conserving", and all work conserving service disciplines have equal average packet delay, provided that the order of serving packets is independent of their lengths. By minimising the number of null polls, exhaustive round robin provides the minimum average packet delay of all schedulers. However, the average delay is not the primary performance criterion for real-time services. Of greater interest is the proportion of packets which arrive within a particular "good service" delay, tg. This is particularly so when a jitter buffer is used by the application. A jitter buffer will discard all packets delayed by more than to, and delay all packets received with lower delay, to ensure that all retained packets have an equal delay of to- Although exhaustive round robin minimises the average packet delay, there is scope to deliver a higher proportion of packets with low to moderate delays. This is the goal of embedded round robin (ERR). Define "busy" stations to be those known to have data to transmit, and "clear" stations to be those which may be idle. The ERR scheduler services clear stations in round robin order. However, between consecutive services of clear stations, it performs an "embedded" round robin cycle of the busy stations. This reduces the time wasted on polling idle or lowrate stations. When there are only a few busy stations, they receive most of the bandwidth, hence their backlog is soon cleared and polling of clear stations is not unduly delayed. However, when the number of busy stations increases, the backlog can remain for some time, and substantially increase the duration of the outer round robin cycle. This

179 Variables: N_max = max. consecutive polls of busy stations busy_count = number of busy stations i := next clear station in RR order Poll station i IF (station i had more data) Mark station i as busy END IF FOR k := 1 to min (N_max, busy_count) j := next busy station in RR order Poll station j IF (station j had no more data) Mark station j as clear END IF END FOR Algorithm 1. One iteration of embedded round robin

could cause packets arriving at "clear" stations to receive unnecessarily poor service. To prevent this, the busy round robin cycle may have to be interrupted to service the next "clear" station. In order to accomplish this, at most Nm^ busy stations are served between each pair of busy stations. Pseudocode for ERR is given as Algorithm 1. A key advantage of ERR over exhaustive RR is that exhaustive RR has no way of knowing which stations other than the current station have data to transmit, whereas ERR does. It is this property which allows the wireless dual queue algorithm17 to detect and combat transient congestion, without causing excessive null polls. 3

Improving High Load Performance

When the number of stations is large, the time taken to cycle through the "clear" list may be excessive if every station in the "busy" list is polled between each pair of polls to clear stations. If the proportion of busy stations remains constant, the cycle time of the clear list grows quadratically with the number of stations. To avoid this, a maximum of iVn^ stations are served from the busy list between polls of the clear list. The maximum time to cycle once through the clear list is then cie + N^f)

(1)

where c is the number of clear stations, e is the polling overhead for sending an empty packet, and / is the time required to transmit a full (maximum length) packet, including the polling time. However the maximum time to cycle once through the busy list becomes (e + iVmax/)6/iVI„ax

(2)

where b is the number of busy stations, compared with bf + ce

(3)

180 for standard round robin. As originally proposed, ERR removes backlogged stations from the main "clear" round robin list when it adds them to the "busy" list. Comparing (2) and (3) shows that when b > iVmaxC, busy stations are actually served less often under ERR than under round robin, rather than more. This condition says that the busy list contains significantly more entries than the clear list, due to network congestion. This phenomenon may be avoided by simply reverting to simple round robin service in this case. However it is exactly at these times that the greater efficiency of ERR is required. An alternative is to leave busy stations in the "main" list when they are added to the busy list. (The term "clear" list is no longer appropriate.) When the busy list is short, this will make minimal difference to the polling order. However, it will ensure that the majority of polls are still sent to busy stations, even when the busy list is long, which improves the network throughput. When a station is removed from the busy list, it should be shifted to the end of the main list, since it is the station with the lowest probability of having data to transmit (zero probability). 4

Least-recently-used ERR

The algorithm hinted at in the previous section can be defined very simply in terms of a single list in which station are sorted in order of how recently they have been polled. This algorithm, given as Algorithm 2, will be called least-recently-used embedded-roundrobin (LRU-ERR). Note that this algorithm is much simpler than Algorithm 1, as the data structures it uses are much simpler. To see how LRU-ERR achieves the objectives set out above, note first that when there are no "busy" stations, it becomes a simple round robin scheduler. This is because the first I F always fails, and the station selected is c a l l s L R U [ 1 ] , the least recently polled station. This is the next station in round robin order. This provides the "outer" round robin loop of ERR. Next, consider the case when the system is lightly loaded, so that every station is polled frequently, and the ELSE I F always fails. The "busy" stations will be polled in LRU (round robin) order, which corresponds to the inner, or embedded, round robin loop of ERR. The round robin polling of "busy" stations will continue until it is interrupted by one of two events. Unless the ELSE I F succeeds, all packets in all queues will be served, causing the "busy" stations one by one to become "clear". When all stations are clear, found will not be set by the FOR loop, causing the final I F to trigger polling of the next "clear" station in round robin order. If the inner round robin cycle always ends this way, then the next station to report m o r e _ d a t a will be the only "busy" station, and thus the inner loop will poll it until it has no more packets. That implies that if t h r e s h is so large that the ELSE I F always fails, then LRU-ERR becomes exhaustive round robin. The ELSE I F prevents LRU-ERR from degenerating to exhaustive round robin by interrupting the inner loop when it is in danger of causing another station to receive bad service. The least recently polled station is in the most imminent danger; if a packet had arrived immediately after its last poll, then it must be polled again within the "good service" time, *G- Moreover, it must be polled in time for all fragments of the original application layer packet to be received before to- For this reason, the threshold used is reduced by a

181 Variables: calls callsLRU[]

= number of current connections = array of calls in order of most recent access, [1] the earliest meanClearTime = mean time required to poll a clear station goodSrvThresh = maximum desired delay margin = "safety margin" (See text.)

thresh := goodSrvThresh - margin found := false FOR i := 1 to calls IF (callsLRU[i] is "busy") found := i BREAK from FOR loop ELSE IF (time since callsLRU[i] polled > thresh) found := 1 BREAK from FOR loop ELSE thresh := thresh - meanClearTime END IF END FOR IF (found = false) found := 1 END IF poll station for callsLRU[i] move callsLRU[i] to end of callsLRU Algorithm 2. One iteration of least-recently-used embedded round robin

margin, tM, which may be tuned to give good performance. In the implementation used here, m e a n C l e a r T i m e was taken to be the time required for a null poll. For that case, the ELSE I F . . . ELSE can be taken outside the loop, since the least recently polled station is the only one which can cause it to succeed. 5

Performance Results

In order to evaluate the performance of LRU-ERR, it was simulated and compared with ERR, round robin (RR) and exhaustive RR. The simulation set up consists of a single cell infrastructure WLAN. The polling mechanism is essentially that defined in the IEEE 802.11 standard2 for the "contention free" mode. Although the standard requires that this mode be punctuated by periods of contention based operation, this was omitted in our simulations. Thirty stations continuously transmit VBR video on the uplink, and the downlink is

182 Table 1. Parameters used in simulations.

Parameter Good service threshold, to Packet expiry timeout, te Time to poll an idle station Time to transmit a maximum length packet (including polling)

value 75 ms 500 ms 0.456 ms 2.83 ms

idle. The lengths of the video frames were taken from the real MPEG traces of several standard video sequences, as used in Reference 18. The 30 sequences were chosen randomly (with replacement) from the 20 available sources, and the starting points within the sequences were chosen randomly. The overall network load was set by scaling the frame sizes by a factor of a, which was significantly less than 1 to model the higher compression needed for wireless transmission. Video frames were segmented into packets of size less than or equal to 2312 bytes, as specified in the 802.11 standard. The raw bit rate was 7.5Mbps. Packets delayed more than te = 500 ms were discarded by the remote stations. The MAC header and inter-frame spacing were set according to the IEEE 802.11 standard2. Other system parameters are given in Table 1. The performance of LRU-ERR was tested under two assumptions for the arrival process of video frames. First, the algorithm will be studied when frames arrive periodically at a rate of 25 frames/sec, after which Poisson arrival will be assumed.

5.1

Periodic frame arrivals

For average loads of a = 0.3 and 0.33, ten independent simulations were performed. Since different video traces were used for each simulation, the performance of each run was substantially different. The packet delays from all runs for each average load were combined and their cumulative distribution functions (CDFs) are shown in Figure 1. For comparison, the results for the same traffic are shown for the original ERR17, standard round robin and exhaustive round robin. When arrivals are periodic with period T, the scheduler can know that the first poll of a station after an interval of T will not be a null poll. Thus there is no throughput benefit associated with having t h r e s h of Algorithm 2 any larger than T. In these experiments the margin, tm, was set such that t h r e s h was marginally above T = 40 ms. The value of A ^ used for which the original ERR was 6, which was empirically found to maximise the proportion of packets received within the good service time. The proportion of packets served by LRU-ERR within the good service threshold was clearly greater than all three other polling strategies. LRU-ERR attempts to maximise the proportion of packets served within time t h r e s h by placing a higher priority on packets about to exceed that time. Because this must increase the delay of the other packets, it is expected that there will be a sharp rise in the CDF immediately before t h r e s h , which is indeed observed.

183 100

K sr •S •D

80

LRU-ERR ERR RR exhaustive RR

60

'"]"' i

.^; -'.,'.-'*

-

a

80

LRU-ERR ERR RR exhaustive RR

•o

60

>. m u

40

1!

•5

8

i

£

I

IB

u-

l'"^r

fj f

100

\JJ%

^?%r--''

i

g.

40

8

20

I

i

20

I

0

0 100 10 packet delay (ms)

1

(a) Relative load a = 0.3.

10 100 packet delay (ms)

(b) Relative load a = 0.33.

Figure 1. Cumulative distribution of packet delays for to = 75 ms with periodic arrivals.

100 £

80

•8

60

LRU-ERR ERR RR exhaustive RR

10 100 packet delay (ms)

(a) Relative load a = 0.3.

LRU-ERR ERR RR exhaustive RR

1

10 100 packet delay (ms)

(b) Relative load a = 0.33.

Figure 2. Cumulative distribution of packet delays for 4Q = 75 ms with Poisson arrivals.

5.2

Poisson arrivals

When video frames arrive according to a Poisson process rather than periodically, the performance of all four scheduling disciplines deteriorates, but exhaustive RR suffers the least. As a result, both exhaustive RR and LRU-ERR perform similarly, and are considerably better than ERR and RR. This can be seen in Figure 2. For these experiments, iVmax = 3 was found to be optimal for ERR. However, it is well know that exhaustive round robin suffers from fairness problems in the presence of a small number of very heavy sources. This case will be considered in the following section.

184 100

2 80 m

?

LRU-ERR ERR RR exhaustive RR

100

E 80

LRU-ERR ERR RR exhaustive RR

7s"

•8 60

60

15 u g. 40

CD

*

10 100 packet delay (ms)

(a) Relative load a = 0.225.

10 100 packet delay (ms)

(b) Relative load a = 0.25.

Figure 3. Cumulative distribution of packet delays for lightly loaded stations, with to = 75 ms, Poisson arrivals and asymmetric loads.

5.3 Asymmetric load In order to test the fairness of LRU-ERR, it was simulated in a network of 30 nodes, of which one produced 10 times the traffic of the others. The delay of packets from the lightly loaded stations was measured and its CDF is shown in Figure 3, when the lightly loaded stations have a = 0.225 and 0.25. Under exhaustive round robin, the heavily loaded station has a dramatic adverse effect on the remaining stations. In contrast, the round robin nature of LRU-ERR distributes bandwidth fairly, as well as efficiently, causing the performance of the majority of stations to be significantly higher than for the other three disciplines. The proportion of packets delivered within the good service time, %G = 0.75 ms, is shown in Figure 4 for a range of loads for both symmetric and asymmetric loads. This again shows that both LRU-ERR and exhaustive round robin effectively eliminate null polls, and consequently give high efficiency for symmetric loads, but that LRU-ERR avoids the unfairness problems inherent to exhaustive round robin.

6

Conclusion

Least-recently-used embedded-round-robin (LRU-ERR) has been presented as an enhancement to the original ERR recently proposed for IEEE 802.11 wireless LANs. This algorithm retains the benefits of ERR demonstrated for light to moderate loads for periodic traffic arrivals, while at the same time improving its performance under high loads and for bursty arrivals. Exhaustive round robin is optimal for avoiding the waste caused by polling idle stations, and for Poisson frame arrivals and symmetric traffic, LRU-ERR achieves performance of the same high standard. However, LRU-ERR outperforms exhaustive round robin under conditions of asymmetric load.

185

LRU-ERR ERR RR exhaustive i i 0.24 0.26

LRU-ERR ERR RR exhaustive i i— 0.28 0.3 relative load, (a)

(a) Symmetric load

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3

relative load, (a)

(b) Asymmetric load

Figure 4. Proportion of packets (from lightly loaded stations) received within t o = 75 ms with Poisson arrivals.

References 1. G. Anastasi, L. Lenzini, E. Mingozzi, "MAC protocols for wideband wireless access: evolution towards wireless ATM," IEEE Personal Commun. , 53-64 (1998). 2. P802.11, Wireless medium access control (MAC) and physical layer (PHY) working group, IEEE 802.11 draft standard, "Wireless medium access control (MAC) and physical layer (PHY) specifications," IEEE Stds. Dept., July 1996, D5. 3. D. Ferrari, D. Verma, "A scheme for real-time channel establishment in wide-area networks," IEEE J. Select. Areas Commun. 8(4), 368-379 (1990). 4. N.R. Figueira, J. Pasquale, "Leave-in-time: a new service discipline for control of real-time communications in a packet-switching network," in Proc. ACM SIGCOMM'95, August 1995, pp. 207-218. 5. D. Verma, H. Zhang, D. Ferrari, "Delay jitter control for real-time communication in a packet switching network," in Proc. IEEE TriCom '91, April 1991, pp. 35-43. 6. A. Demers, S. Keshav, S. Shenker, "Analysis and simulation of a fair queueing algorithm," in Proc. ACM SIGCOMM 89, Sept. 1989, vol. 19, no. 4, pp. 1-12. 7. A.K. Parekh, R.G. Gallager, "A generalized processor sharing approach to flow control in integrated services networks: the single-node case," IEEE/ACM Trans. Networking 1(3), 344-357 (1993). 8. S.J. Golestani, "A self-clocked fair queuing scheme for broadband applications," in Proc. IEEEINFOCOM '94, 1994, pp. 636-646. 9. D. Saha, S. Mukherjee, S.K. Tripathi, "Carry-over round robin: a simple cell scheduling mechanism for ATM networks," IEEE/ACM Trans. Networking 6(6), 779-796 (1998). 10. M. Shreedhar, G. Varghese, "Efficient fair queueing using deficit round robin," IEEE/ACM Trans. Networking 4(3), 375-385 (1996) 11. D.A. Hayes, M. Rumsewicz, and L.L.H. Andrew, "Quality of service driven packet scheduling disciplines for real-time applications: looking beyond fairness," in Proc. IEEEINFOCOM'99, 1999, vol. 1, pp. 405^12.

186 12. J. Nagle, "On packet switches with infinite storage," IEEE Trans. Commun. COM35, 435-438 (1987). 13. N. Movahhedinia, G. Stamatelos, H.M. Hafez, "Polling-based multiple access for indoor broadband wireless systems", in Proc. PIMRC 95, 1995, vol. 3, pp. 10521056. 14. R. Kautz, A. Leon-Garcia, "A Distributed self-clocked queueing architecture for wireless ATM networks," in Proc. PIMRC'97, 1997, vol. 1.1, pp. 189-193. 15. C.-S. Chang, K.-C. Chen, "Guaranteed quality-of-service wireless medium access by packet-by-packet generalized processor sharing algorithm," in Proc. IEEE Int. Conf. Commun. (ICC)'98, June 1998, vol. 1, pp. 493^197. 16. R.S. Ranasinghe, D. Everitt, L.L.H. Andrew, "Fair Queueing scheduler for IEEE 802.11 based wireless multimedia networks," Multiaccess, Mobility and Teletraffic in Wireless Communications (MMT'99), Oct. 1999, vol. 4, pp. 241-251. 17. R.S. Ranasinghe, L.L.H. Andrew, D.A. Hayes and D. Everitt, "Scheduling disciplines for multimedia WLANs: Embedded round robin and wireless dual queue," in Proc. IEEE Int. Conf. Commun. (ICC) '01, Helsinki, Finland, June 2001. 18. O. Rose, "Statistical properties of MPEG video traffic and their impact on traffic modeling in ATM systems," in Proc. 20th Annual Conf. Local Comp. Network, Minneapolis, CA, Oct. 1995. pp. 397^06.

A NOVEL MAC PROTOCOL FOR POWER EFFICIENT SHORT-RANGE WIRELESS NETWORKING

T. Y. CHUI AND W. G. SCANLON Centre for Communications Engineering, School of Electrical & Mechanical Engineering, University of Ulster, Shore Road, Newtownabbey, Co. Antrim, Northern Ireland, BT37 OQB. E-mail: [email protected],

[email protected]

A novel wireless medium access control protocol is presented. With the use of resynchronisation and scheduling pointers, the proposed protocol offers improvements in the power efficiency of portable terminals and is robust in the presence of difficult radio propagation conditions. To quantitatively determine the efficiency gains the steady-state performance of the new PDAnet protocol was compared with standard Bluetooth technology. The results demonstrate a significant improvement in terminal energy efficiency using the new approach when more than 1 slave device is active within a short-range wireless network (piconet). Depending on the packet length and number of active slaves, the amount of energy wasted by the needless activation of a terminal's RF circuitry was reduced by up to 70.8 %. In addition, PDAnet could deliver up to 42.8 % more data for equal energy consumption than Bluetooth, while throughput was not significantly affected. Further improvements in network performance can also be achieved by taking advantage of the PDAnet pointer-based approach and introducing packet-by-packet adaptive scheduling.

1. Introduction Medium access control (MAC) protocols are extremely important in the design of short-range wireless network systems. Particular physical layer problems include intermittent hidden terminals, statistically varying link quality [1], and the energy constraints of portable terminals [2]. These issues are partially addressed by using packet-radio techniques. For example, personal area network (PAN) systems such as Bluetooth [3] use time division techniques to reduce the energy consumption of terminals. The RF transceiver is only operational when transmitting, receiving a correctly addressed packet, or inspecting the header of each packet transmitted by the master node in the network. This approach is reasonably energy efficient for small numbers of nodes. However, the energy consumed during the inspection of packets destined for other nodes is wasted. PDAnet is a new short-range wireless networking system specifically designed for hand portable computers and consumer devices such as personal digital assistants. The main aims of PDAnet are to reduce terminal energy consumption and

187

188 to implement intelligent, link-adaptive packet scheduling to make efficient use of the radio environment [4]. This paper introduces novel aspects of the PDAnet MAC protocol that offer reduced energy consumption for active terminals in a heavily loaded piconet during active mode. The energy efficiency of the MAC protocol is quantitatively compared with that of Bluetooth. 2. PDAnet MAC Protocol The PDAnet MAC protocol is based on a master-slave system with a superframe (polling cycle) that is divided into a fixed number of equal time slots. PDAnet can be implemented as a modification to the existing Bluetooth MAC with minimal changes, or it can form the basis of a proprietary wireless network. For example, the PDAnet downlink packet format shown in Figure 1 is identical to Bluetooth but with the addition of one bit (Master/Slave indication) at the header and a pointer field (18 bits). Uplink packets (Figure 2) do not require pointers and therefore only the Master/Slave indication bit is added. Access Code (68-72 bits)

Header (57 bits)

_,—-"'

AM.ADDR (3 bits)

Packet Type (4 bits)

Flow (lbit)

I

ARQN SEQN M/S Header Error Check (HEC) (8 bits) (1 bit) (lbit) (1 bit)

(0-2728 bits) V

I

FEC 1/3 encoding replicates data 3 times

Payload

Pointers (18 bits)

»

V

FEC 2/3 encoding Resynchronisation Pointer (6 bits)

""*>,, Next Slot Pointer (6 bits)

Figure 1. PDAnet MAC downlink packet format. Access Code (68-72 bits)

^^.~~~ ' - " " AM_ADDR (3 bits)

Packet Type (4 bits)

Header (57 bits)

Payload (0-2744 bits)

FEC 1/3 encoding replicates data 3 times Flow ARQN SEQN M/S Header Error Check (HEC) (8 bits) (lbit) (lbit) (lbit) (1 bit)

Figure 2. PDAnet MAC uplink packet format. The transmission mode is time division duplex and the master embeds a resynchronisation pointer (RP) in the header of each downlink (master-slave) packet. The RP indicates the number of remaining time slots before the transmission

189

of a scheduled piconet broadcast beacon. The beacon contains a range of information including slot scheduling data, the master's ID and the services currently available in the piconet. In frequency hopping systems, the RP could also indicate the transmission channel for the beacon. The slot following a broadcast beacon is designated for carrier sense multiple access (CSMA) (Figure 3). Unsynchronised slaves can then contend in this slot to request service from the master. Feedback from successful requests is included in the next broadcast beacon. One Polling Cycle = 64 slots (6 bits)

's'ir Registration Beacon

k- s ™^i CSMA Contention Slot

Mas Sla

en

Master Slot

Slave n to Master

Mas Sla

Slave Slot

Slave n to Master

Master to

Slave n to Master

Registration Beacon

CSMA Contention Slot

4 Master Slot

Slave Slot

RP=1 RP=59 RP=6i

Figure 3. PDAnet MAC superframe operation illustrating the resynchronisation pointer. Figure 4 shows how a new slave can synchronise with the piconet. The slave activates its receiver and listens to any downlink slot, extracting the RP from the packet header before entering low-power sleep mode until the next broadcast beacon. After listening to the broadcast beacon, the slave can participate in the CSMA session if desired. It transmits its terminal ID and connection request during the CSMA period. The slave then sleeps until the next scheduled broadcast beacon. If the slave was successful it will find its terminal ID and initial allocated data transmission slot during the broadcast. Otherwise, the terminal repeats the access procedure again. This approach avoids the unnecessary latency associated with systems such as Bluetooth where the master periodically suspends the piconet activity to perform a lengthy inquiry, page and scan. Although the RP is an additional link layer overhead, it allows fast, controlled access to the piconet. In addition, the RP provides a method of rapid resynchronisation for any terminal that has recently lost connection with the network because of propagation related packet loss or any other reason. Furthermore, idle terminals can also use the contention period to signal the master and request a change of operating mode. PDAnet implements packet scheduling using a next slot pointer (NSP). A common approach in wireless networks is to pre-allocate a block of slots for each slave within each superframe [5]. However, fast changing radio propagation

190 conditions may lead to sustained packet loss during the pre-allocated block and network throughput will be reduced. In PDAnet, the NSP introduces a higher level of control and enables the master to intelligently schedule slots on a packet-bypacket basis. Slaves normally transmit immediately following receipt of a downlink packet. The NSP, extracted from the header of the last downlink packet received, then indicates when to expect the next downlink. The slave simply enters sleep mode until that time. If a poor quality link is detected, the master can react quickly within the current polling cycle by reducing any further allocation of slots to the relevant slave. Therefore, wasted slots are kept to minimum. The scheduling information in the broadcast beacon consists only of the next allocated slot for each slave, rather than the next batch of allocated slots. The broadcast overhead is therefore reduced.

Listen to any downlink transmission

Send ID in CSMA contention period

X

Extract RP from packet header

•

Enter low power sleep mode

zz

Enter low-power sleep mode

31

Wake up after (x) slots Listen to broadcast beacon

X

(a)

Decide to join / resynchronise

(b)

Figure 4. PDAnet (a) synchronisation procedure, (b) registration procedure. In popular applications such as a campus Intranet, the portable terminals are usually clients who request a large amount of data from fixed network servers via access points (i.e. a piconet master). As the data flow may be asymmetric, PDAnet gives the master control over the ratio of downlink to uplink slots. It is also possible to extend this adaptability by implementing an intelligent scheduler at the master based on the average data loading of each terminal over the past few superframes. Similarly, if required the master can take into account the link conditions by using the packet error rate or dropped packet rate as a factor in the scheduling algorithm [6]. This approach will also maximise network throughput and reduce overall energy consumption. Sending packets over a poor quality link or turning on a RF transceiver for a null packet will increase terminal energy consumption.

191 3. Bluetooth Overview The Bluetooth MAC uses time slots of 625 \as duration and devices can operated in 4 different transmission modes: active, sniff, hold and park mode. There are 3 different packet sizes for active mode data transmission, each single packet occupying 1, 3 or 5 slots. The Bluetooth physical layer uses 1 Msymbol/s Gaussianfiltered frequency shift keying (GFSK) modulation in a 2.45 GHz frequency hopping spread spectrum (FHSS) system. The maximum transmission duration in a single slot is limited to 366 symbols due to the need for a guard period to accommodate frequency hopping. Therefore, to achieve the maximum asymmetric downlink throughput of 723.2 kbps, frequency hopping does not occur during the 5slot packet. Although a higher throughput is obtained, there is a greater risk of packet collisions or interference caused by either another Bluetooth piconet or ISMband user in a close proximity. Increasing the packet length also increases the packet error rate for a particular level of link quality (bit error rate) and the number of packet retransmissions will grow rapidly. In contrast to PDAnet, Bluetooth employs complex procedures to establish connections with relatively long access times. The Page and Scan procedure can require up to 5 seconds to complete when the device type is known. Each Bluetooth slave must listen to the header of every downlink packet. Therefore, if there are 7 active terminals in a piconet, 6 terminals are listening to each downlink packet header unnecessarily. However, if a 5-slot packet is detected the other slaves will enter sleep mode until the next downlink slot. The maximum sleep duration in PDAnet is 2 slots less than the superframe length as all active slaves must wake and listen to the broadcast beacon. 4. Energy Consumption Results A steady-state analysis was carried out to determine how the proportion of energy wasted is related to the number of active slaves in both PDAnet and Bluetooth systems. For a fair comparison, the analysis only considered the 3 packet lengths available using Bluetooth and both systems were assumed to operate over the same hardware. The relevant dynamic connection procedures were not included and each slave is assumed to be 100 % loaded (data always available for both uplink and downlink). Tables 1 to 3 compare the calculated normalised energy consumption figures, throughput and transmission efficiency (kbits/mJ) for both Bluetooth and PDAnet. The normalised energy consumption figures are based on RF hardware consuming 40 mJ/s when receiving, 32 mJ/s when transmitting and 0.065 mJ/s when idle. The analysis is based on a superframe of 64 slots and takes into account the relevant system overheads (the PDAnet broadcast beacon, additional PDAnet pointers and the inspection of packet headers in Bluetooth). The analysis is based on an error free environment therefore only unprotected packet types are being used.

192 Number of active slaves 1 2 3 4 5 6 7

Normalised energy consumption (mJ/s) 36.0 20.0 14.7 12.1 10.5 9.4 8.6

Downlink throughput per slave (kbps) 172.8 86.4 57.6 43.2 34.6 28.8 24.7

Uplink throughput per slave (kbps) 172.8 86.4 57.6 43.2 34.6 28.8 24.7

Total throughput (kbps) 345.6 172.8 115.2 86.4 69.2 57.6 49.4

Transmission efficiency (kbits/mJ) 9.60 8.62 7.82 7.16 6.60 6.12 5.71

Table la. Steady-state results for 1 slot unprotected packet in Bluetooth. Number of active slaves 1 2 3 4 5 6 7

Normalised energy consumption (mJ/s) 36.1 18.4 12.5 9.5 7.8 6.6 5.7

Downlink throughput per slave (kbps) 157.5 78.8 52.5 39.4 31.5 26.3 22.5

Uplink throughput per slave (kbps) 170.1 85.1 56.7 42.5 34.0 28.4 24.3

Total throughput (kbps) 327.6 163.9 109.2 81.9 65.5 54.7 46.8

Transmission efficiency (kbits/mJ) 9.08 8.91 8.75 8.59 8.44 8.29 8.15

Table lb. Steady-state results for 1 slot unprotected packet in PDAnet. Number of active slaves 1 2 3 4 5 6 7

Normalised energy consumption (mJ/s) 38.0 20.0 14.1 11.1 9.3 8.1 7.2

Downlink throughput per slave (kbps) 585.6 292.8 195.2 146.4 117.1 97.6 83.7

Uplink throughput per slave (kbps) 86.4 43.2 28.8 21.6 17.3 14.4 12.3

Total throughput (kbps) 672.0 336.0 224.0 168.0 134.4 112.0 96.0

Transmission efficiency (kbits/mJ) 17.68 16.77 15.94 15.19 14.51 13.89 13.32

Table 2a. Steady-state results for 3-slot unprotected packet in Bluetooth. Number of active slaves 1 2 3 4 5 6 7

Normalised energy consumption (mJ/s) 38.0 19.4 13.1 10.0 8.2 6.9 6.0

Downlink throughput per slave (kbps) 570.2 285.1 190.1 142.5 114.0 95.0 81.5

Uplink throughput per slave (kbps) 85.1 42.5 28.4 21.3 17.0 14.2 12.2

Total throughput (kbps) 655.3 327.6 218.5 163.8 131.0 109.2 93.7

Transmission efficiency (kbits/mJ) 17.23 16.92 16.63 16.34 16.06 15.80 15.54

Table 2b. Steady-state results for 3-slot unprotected packet in PDAnet.

193 Number of active slaves 1 2 3 4 5 6 7

Normalised energy consumption (mJ/s) 38.7 20.0 13.8 10.7 8.9 7.6 6.7

Downlink throughput per slave (kbps) 723.2 361.6 241.1 180.8 144.6 120.5 103.3

Uplink throughput per slave (kbps) 57.6 28.8 19.2 14.4 11.5 9.6 8.2

Total throughput (kbps) 780.8 390.4 260.3 195.2 156.1 130.1 111.5

Transmission efficiency (kbits/mJ) 20.19 19.48 18.82 18.21 17.63 17.09 16.57

Table 3a. Steady-state results for 5-slot unprotected packet in Bluetooth. Number of active slaves 1 2 3 4 5 6 7

Normalised energy consumption (mJ/s) 38.7 19.7 13.4 10.2 8.3 7.0 6.1

Downlink throughput per slave (kbps) 707.7 353.9 235.9 176.9 141.5 118.0 101.1

Uplink throughput per slave (kbps) 56.7 28.4 18.9 14.2 11.3 9.5 8.1

Total throughput (kbps) 764.4 382.3 254.8 191.1 152.8 127.5 109.2

Transmission efficiency (kbits/mJ) 19.76 19.41 19.08 18.76 18.44 18.14 17.85

Table 3b. Steady-state results for 5-slot unprotected packet in PDAnet. Figure 5 was derived from Tables 1 to 3 and confirms that PDAnet wastes less energy than Bluetooth when more than two slaves are active. With PDAnet, the amount of wasted energy does not depend on the packet size because the overheads (broadcast and next slot pointer) are fixed. When there are 7 active slaves in the network and 5 slot packets are being used, 17.9 % of the total energy consumption is wasted by Bluetooth, falling to 11.1 % for PDAnet. This is an improvement of 38 %. However, when using a 1-slot packet, the default packet type in most implementations, 40.5 % of the total energy is wasted in Bluetooth, falling to 11.8 % with PDAnet. Here the use of the RP and NSP gives a 70.8 % improvement over Bluetooth The pointers and broadcast beacon used in PDAnet marginally reduce the downlink data throughput (by 5.2 % for a 1-slot packet, falling to 2.1 % for a 5-slot packet). However, the PDAnet transmission efficiency (amount of data bits transmitted per unit energy consumed) is still higher than with Bluetooth (Figure 6). To demonstrate the significance of these results, Figure 7 shows the improvement in transmission efficiency offered by PDAnet which can be up to 42.8 % (for 7 active slaves and using 1-slot packets). Higher transmission efficiency means longer battery lifetime for the mobile terminal.

194 45

Tr=========;------------------, -+-1 slot (Bluetooth) ___ 1 slot (PDAnet)

40

-.- 3 slots (Bluetooth) 3 slots (PDAnet) -lIE- 5 slots (Bluetooth)

-+- 5 slots (PDAnet)

OlF"-----r----~----~----~----~---_____1

1

2

3

5

4

7

6

Number of active slaves

Figure 5. Ratio of wasted energy for different packet lengths. 25,------------------.----=--------=:--,------, -+-1 slot (Blueloolh)

I

-.-3 slots (Bluetooth) ~ 5 slots

(Bluetooth)

--- 1 slot(POAnet)

:.+-3 slots (PDAnet)

....... 5 slots (PDAnel)

20t::~.~=_ . ~~~~

E 15 [====::::::::=~~=::;;:::::;:::=:==~;;;==~=:=:==== ----......

.

~ --.

i

~

~ 10 " ' t - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 1

5C==:·····----~····=····-==······=·····-~·····;;······-~·····d

0+-----r-----~---~----___,.----~-------1

1

2

3

4

5

6

7

Number 01 active slaves

Figure 6. Transmission efficiency (data transmitted per energy consumed).

The results presented in Tables 1 - 3 and Figures 5 - 7 suggest that the PDAnet approach is beneficial for mobile terminals when the piconet is heavily loaded. In cases where the data loading in the lower there are currently a number of power

195

saving schemes available. These include sniff mode in Bluetooth or power saving polling (PSP) mode in IEEE 802.11 and they provide reasonable power saving for handheld terminals [7].

3

4

5

Number of active slaves

Figure 7. Relative improvement in transmission efficiency provided by PDAnet. However, the PDAnet will also deliver power savings in light loading conditions by means of the regular broadcast beacon. For example, a terminal may periodically listen to the broadcast beacon once every polling cycle to find out if there is data awaiting to be downloaded from the master. To extend power saving even further, a handheld terminal can listen the broadcast beacon every so many polling cycles. However, there is a direct trade-off between reasonable latency (on the downlink) and the desired power saving. The proposed PDAnet protocol is not only concerned with terminal power savings as it also introduces a high degree of flexibility in setting the defer interval between data transmissions. The use of packet embedded pointers enables flexible scheduling without the need to exchange higher layer messages to negotiate for an agreed sniff period. 5. Conclusion Details of the MAC layer protocol of a new short-range wireless networking system were presented. The protocol was designed for optimal performance with hand-held computing devices operating in variable channel conditions. Packet-by-packet scheduling and the use of a resynchronisation pointer enable rapid recovery from

196 packet loss, while providing a fast method of terminal registration. One of the most important aspects of the system, however, is the improved energy efficiency for mobile terminals. Results were presented that demonstrate how the new system is more energy efficient than Bluetooth when there is more than one active slave in a piconet. The use of packet embedded pointers also enables the use of intelligent slotby-slot scheduling algorithms to optimise the network and terminal resources. Future study will use discrete event modelling to analyse die effect of various scheduling algorithms, and to assess the resulting network performance in the presence of both channel and load variability.

6. Acknowledgement This work is supported by Nortel Networks UK and the Northern Ireland Industrial Research and Technology Unit (IRTU). References 1. Villanese, F., Scanlon, W.G. & Evans, N.E., Statistical characteristics of pedestrian-induced fading for a narrowband 2.45 GHz indoor channel. Proc. IEEE Conf. Vehicular Technology, (2000), pp. 745-751. 2. Bhagwat, P., Bisdikian, C , Korpeoglu, I., Krishma, A., and Naghshineh, M., System design issues for low-power, low-cost short range wireless networking. IEEE Intl. Conf. Personal Wireless Communications (1999) pp. 264-268. 3. Bluetooth specification, www.Bluetooth.com. 4. Bhagwat, P., Bhattacharya, P., Krishna, A. and Tripathi, S.K., Using channel state dependent packet scheduling to improve throughput over wireless LANs. Wireless Networks (1997) Vol. 3, No.l pp. 91-102. 5. Inoue, M., Channel state dependent resource scheduling for wireless message transport with framed ALOHA-reservation access protocol. 1EICE Trans. Fundamentals of Electronics, Communications and Computer Sciences (2000) Vol. E83-A, No.7, July, pp. 1338-1346. 6. Anvekar, D.K. and Kapoor, M., Frequency look ahead and link state history based interference avoidance in wireless pico-cellular networks. The Proceedings of IEEE International Conference on Personal Wireless Communications (ICPWC 2000) (2000), pp. 434-438. 7. Bing, B., High-speed wireless ATM and LANS (Artech House Inc., 2000) pp. 100-102.

ROC: A WIRELESS MAC PROTOCOL FOR SOLVING THE MOVING TERMINAL PROBLEM Chi-Hsiang Yeh Dept. of Electrical and Computer Engineering Queen's University Kingston, Ontario, K7L 3N6, Canada E-mail: [email protected] In this paper, we propose a distributed quality-of-service (QoS)-guaranteed medium access control (MAC) protocol, called the request-to-send/object-to-send/clear-to-send (ROC) protocol, by adding an object-to-send (OTS) mechanism to the request-to-send/clear-to-send (RTS/CTS)based protocols. We identify the moving terminal problem for mobile ad hoc networks, and present its solutions based on ROC and several ad hoc protection techniques. We show that ROC and the associated techniques can solve the moving terminal problem, guaranteeing QoS in high-mobility ad hoc networks. ROC can also be extended to obtain MAC protocols with variable transmission radii, which may considerably increase network throughput and reduce power consumption.

1

Introduction

IEEE 802. I I 8 defines the MAC and physical layer standards for the license-free industrial, scientific, and medical (ISM) bands 5 allocated by the Federal Communications Commission (FCC) in the U.S. The IEEE 802.11 network architectures include wireless LANs n , where there is at least one or several access points, and ad hoc networks 1,3 , where there is no infrastructure - both of which are expected to become crucial to the future networking environments, and have received tremendous attentions in the past few years. The MAC protocol of IEEE 802.11 is a carrier-sense multiple access protocol with collision avoidance (CSMA/CA) 1 ' 6,11 , which supports RTS/CTS operations 9 on an optional basis. Request-to-send/clear-to-send (RTS/CTS)-based protocols 2>6>9>10 are distributed MAC protocols that can solve the hidden terminal problem and the exposed terminal problem in ad hoc networks. In an RTS/CTS-based protocol, a transmitter first sends an RTS message to acquire the "floor" for later transmission, and a receiver replies with a CTS message if the channel is clear. Previous RTS/CTS-based protocols, however, cannot solve the moving terminal problem, where data packets may still collide at receivers after their RTS/CTS dialogues and the reservations made may conflict with each other at a later time due to movements of transmitters and/or receivers. In this paper, we propose the request-to-senaVobject-to-senaVclear-to-send(ROC) protocol, by adding an object-to-send (OTS) mechanism and several associated features and mechanisms to the previous RTS/CTS scheme 2,6 ' 9 ' 10 . In an ROC-based protocol, a transmitter first multicasts an RTS message to all stations (i.e., mobile hosts 197

198 (MHs) and/or base stations) within its interference range, or protection range (i.e., a range somewhat larger than its interference range for protection purposes against mobility), to request for approval. If a station is against the request, it replies the intended transmitter with an OTS message to block the intended transmission. We formulate the moving terminal problem that is unique for mobile ad hoc networks and wireless LANs or home networks with mobile devices, and show through the method of exhaustion that ROC, Geographical Reservation And Clusterhead Election (GRACE), cell protection, geographical area protection, and extension protection can solve the moving terminal problem, guaranteeing QoS with high probabilities. ROC is also suitable for wireless networks that have variable transmission radii, which cannot be efficiently supported by MAC protocols based on busy tone, and can support distributed assignments of codes for code division multiple access (CDMA). 2

The Request-to-send/Object-to-send/CIear-to-send (ROC) Protocol

In this section we propose the request-to-senaVobject-to-send/clear-to-send(ROC)protocol, a new distributed MAC protocol for QoS guarantees in mobile ad hoc networks. ROC is derived by adding an object-to-send(OTS) mechanism and several associated features and mechanisms to the previous RTS/CTS scheme. In ROC, a transmitter first multicasts an RTS message to all stations (i.e., MHs and/or base stations) within its interference range or an enlarged region to be referred as its protection range, rather than its coverage range only. The purposes of RTS messages in ROC are (1) to inquire the receiver whether the interference at its current location (and possibly at its predicted future locations) is low enough to receive its packet, possibly through carrier sensing on top of the RTS messages it has received (and whether there are scheduled transmissions that will be interfered by the intended transmission), and (2) to inquire other stations within its interference/protection range whether the intended transmission will collide with the packets that they are receiving or will receive. Note that reservations for scheduled transmissions will be made possible by our ROC scheme, as will be explained later. Note also that if the interference/protection range of an intended transmission is greater than the maximum transmission radius of the transmitter, we need to use a relayed geocasting mechanism to forward the RTS message to all the stations within its interference/protection range. If a station that received the RTS message is receiving a packet or will be receiving a packet during an overlapped period of time, it informs the intended transmitter with an OTS message, possibly after relaying, and the intended transmitter has to backoff and request to send again at a later time. On the other hand, if the intended transmitter receives a CTS from the receiver and does not receive any OTS messages (after waiting for a sufficiently long period of time), it can then safely start its transmission. Note that the additional OTS mechanism does not require any additional hardware as compared to RTS/CTS-based protocols (such as an additional transceiver required for busy tone 12 ). The reason ROC-based protocols can provision QoS guarantees is that their OTS mechanism can enable the "reservation" and "blocking" of radio resources periodically along the routing path of a QoS session. More precisely, when other stations in-

199

tend to transmit packets that will interfere with a scheduled packet (that has reserved the radio resources), the receiver of the scheduled packet will send OTS messages to block the intended transmissions. Also, if a CTS message received by the transmitter of a scheduled packet (that has reserved the resources) indicates that the intended receiver plans to receive a packet that will interfere with the scheduled packet, the transmitter of the scheduled packet will also send an OTS message to the intended receiver and/or directly to the intended transmitter to block the transmission. Thus, the OTS and associated mechanisms of ROC in effect reserve and block the required resources along the routing paths of QoS sessions in a wireless network. ROC leads to distributed MAC protocols that can guarantee QoS in wireless networks, rather than improving the service quality only by using priority-based techniques as in the current versions of IEEE 802.11 8 and HIPERLAN4 and in several other previous MAC protocols. There are several challenging issues associated with the moving terminal problem that considerably complicates the required mechanisms for providing QoS guarantees in ad hoc wireless mobile networks. In particular, when QoS sessions reserve resources through moving MHs for a sufficiently long time, which we expect to be commonplace in ad hoc networks, wireless LANs with relaying capability, as well as new wireless network architectures such as dynamic cellular networks 15 , the moving terminal problem may become especially severe. For example, parts of the paths for two QoS sessions may move too close due to the mobility of their relaying MHs so that the "reserved resources" may then conflict in both time and space. As a result, a receiver belonging to one of the two QoS sessions may send OTS messages to block the transmissions of the other QoS session, and vice versa, so that QoS may not be guaranteed anymore. To solve the preceding problem, which we refer to as mutual objection of the moving terminal problem, we require several accompanying mechanisms for the ROC scheme, including a rescheduling algorithm and a geographical reservation mechanism such as GRACE17, where resources are reserved for backbone nodes or clusterheads, taking into account the possibility that they may be moving within certain geographical regions. An adaptable reservation mechanism such as the penaltybased adaptable reservation mechanism 16 or the Priority-based Aggressive Reservation Interactive Scheme (PARIS), can then be incorporated into ROC as MAC/DA was applied to reservation-enabled MAC protocols in 16 since ROC can support a reservation mechanism.

3

The Moving Terminal Problem and Its Solutions Based on ROC

In this section, we exhaustively describe all possible scenarios of the moving terminal problem in ad hoc mobile wireless networks, and then present solutions based on ROC as well as other associated mechanisms such as GRACE 17 and several ad hoc protection techniques.

200

(a)

(b)

Figure 1: (A) An on-going transmitter moves from location 7j to location T2 so that it interferences with the reception at MH Rx, where the dark circle represents its coverage range and the gray circle represents its interference range. (B) An on-going receiver moves from location Dj to location D\. Since MH D\ moves away from its upstream transmitter Su MH Si may have to increase its power level, so that the on-going transmission interferences with the reception at MH R2.

3.1 Interfering with Transmitting MHs In this subsection we consider the cases where interference occurs dee to on-going transmissions or intended transmissions that have completed their ETS/CTS or ETS/OTS/CTS dialogues. In the following subsections we will then only consider the cases where interference occurs involving scheduled QoS transmissions. Collisions Caused by Moving Transmitters A moving on-goieg transmitter may move so that its interference range covers a nearby on-going receiver (see Fig. la). The reception at the nearby on-going receiver may then be collided. To avoid such collisions, we simply use protection ranges somewhat larger than the corresponding interference ranges and multicast RTS or CTS messages to all MHs with such a protection range before a transmission or reception. By limiting the maximum size of packets, the increases in the protection ranges over the corresponding interference ranges are also limited so that the wasted resources are negligible. Collisions Caused by Moving Receivers A moving on-going receiver may move into the interference range of a nearby on-going transmitter. The reception at the moving on-going receiver may then be collided. A moving on-going receiver may also move farther away from its upstream transmitter. If the power level of the transmitter is increased, the interference range associ-

201

Ri

(a)

(b)

Figure 2: (A) A potential transmitter moves from location 7j to location T2 so that its potential interference range covers MH R\. MH T2 may not be aware of the on-going or scheduled receptions at MH R\. (B) MH T2 multicasts an RTS message to all the MHs within its interference range through a control channel. MH R\ replies with an OTS message to block the intended transmission.

ated with its transmission may increase so that the new interference range may cover a nearby on-going receiver. The reception at the nearby on-going receiver may then be collided (see Fig. lb for an example). Similar to the collisions caused by moving transmitters, such collisions can all be avoided by using protection ranges somewhat larger than the corresponding interference ranges. Unknown Transmissions Caused by Moving Intended Transmitters A moving potential transmitter may move so that its potential interference range covers a nearby ongoing receiver (see Fig. 2a). The moving potential transmitter may not be aware of an on-going transmission (if carrier sensing is not used) or a future transmission which has completed its RTS/CTS dialogue. If the moving potential transmitter intends to transmit packets, it will multicast an RTS message to nearby MHs. A nearby MH that is, or will be, receiving a packet can then send an OTS message to block the intended transmission, avoiding possible collision (see Fig. 2b). In addition to the OTS mechanism, such problems can also be avoided by using protection ranges somewhat larger than the corresponding interference ranges. Unknown Transmissions Caused by Moving Intended Receivers A moving potential receiver may move into the interference range of a nearby on-going receiver. The moving potential receiver may not be aware of an on-going transmission (if carrier sensing is not used) or a future transmission which has completed its RTS/CTS dialogue. If the moving potential receiver intends to receive packets, it will multicast an CTS message

202

to nearby MHs. A nearby MH that is, or will be, transmitting a packet can then send an OTS message to block the intended transmission, avoiding possible collision. In addition to the OTS mechanism, such problems can also be avoided by using protection ranges somewhat larger than the corresponding interference ranges. 3.2 Moving Non-QoS MHs Interfering with Static QoS MHs In this subsection, we present the scenarios when an MH without QoS-guaranteed traffic (to be referred to as a non-QoS MH), either a transmitter or a receiver, moves too close to an MH with QoS-guaranteed traffic (to be referred to as a QoS MH), either a transmitter or a receiver. Moving Non-QoS Transmitters When a non-QoS potential transmitter moves so that its new interference range covers one or several QoS receivers, it may not know the schedules of those QoS receivers. Therefore, the non-QoS transmitter may intend to transmit packets at times somewhat earlier than scheduled QoS transmissions, and multicast its RTS messages to MHs within its interference/protection range. To avoid collision, the QoS receivers with a conflict schedule can then send an OTS message to the non-QoS intended transmitter to block its transmissions. Moving Non-QoS Receivers When a non-QoS potential receiver moves into the interference range of one or several QoS transmitter, it may not be aware the schedules of those QoS transmitters. Therefore, when its upstream non-QoS transmitter send an RTS message to it, the non-QoS potential receiver may reply with a CTS message. This CTS may be multicast to MHs whose interference ranges cover the non-QoS potential receiver, and a QoS transmitter with a conflict schedule at a somewhat later time can then send an OTS message to the non-QoS intended receiver and/or its upstream non-QoS intended transmitter to block the transmissions. Note that the above mechanism can avoid collision and improve efficiency, but is optional in ROC since QoS transmissions will not be collided without such a mechanism. With such a mechanism, however, ROC can become collision-free for data packets. 3.3 A Moving QoS MH Interfering with Static Non-QoS MHs In this subsection, we present the scenarios when a QoS MH, either a transmitter or a receiver, moves too close to a non-QoS MH, either a transmitter or a receiver. A Moving QoS Transmitter When a QoS transmitter moves so that its new interference range covers one or several non-QoS receivers, these non-QoS receivers may not know the schedules of the QoS transmitter. Therefore, such a non-QoS receiver may intend to receive packets at times somewhat earlier than scheduled QoS transmissions, and multicast its CTS messages to MHs whose interference/protection ranges cover it. To keep the scheduled slots for QoS transmissions, the QoS transmitter will then send

203

Figure 3: An MH on a QoS path moves from Ef2 to D2 so that the transmission radius of its upstream MH T2 has to be increased. The new interference range of MH T2 cover several MHs on another QoS path, so that QoS cannot continue to be guaranteed if their schedules coniict and cannot reschedule it successfully.

an OTS message to the non-QoS intended receiver or directly to its upstream non-QoS intended transmitter to Mock the transmissions with a conflict schedule. This mechanism is mandatory in ROC in order to guarantee QoS. A Moving QoS Receiver When a QoS receiver moves into the interference range of one or several non-QoS intended transmitters, these non-QoS intended transmitters may not be aware the schedules of the QoS receiver. Therefore, such a non-QoS intended transmitter may intend to send packets at times somewhat earlier than scheduled QoS transmissions, and multicast its RTS messages to MHs within its interference/protection range. To keep the scheduled slots for QoS transmissions, the QoS receiver will then send an OTS message to the non-QoS intended transmitter to Mock the transmissions with a conflict schedule. This mechanism is mandatory in ROC in order to guarantee QoS. 3.4

Moving Intermediate QoS MHs Interfering with Each other

In this subsection, we present the scenarios when an MH with QoS-guaranteed traffic, either a transmitter or a receiver, moves too close to another MH with QoS-guaranteed traffic, either a transmitter or a receiver. When one or several QoS MHs move so that at least two QoS MHs with conflict schedules become to close to each other, then the original schedule cannot both be enforced (see Fig. 3 for an example). We will then have to reschedule the QoS slots, or

204

rely on QoS adaptation to reduce the required QoS slots when rescheduling cannot resolve the conflict (e.g., when the sum of their required bandwidth exceeds the available bandwidth at that location), or when rescheduling cannot be finished in time. To avoid, or at least considerably reduce, the occurrence of such scenarios, we may rely on Geographical Reservation And ClusterheadElection (GRACE) combined with an ROC-based reservation mechanism. In GRACE, the physical space is partitioned into cells (see Fig. 4) with any appropriate shapes and sizes, according to the network topology, the available geographical information, the traffic conditions, etc. It is allowed to have multiple cells that overlap at a location. A clusterhead is elected from each of the cells, serving as a resource broker and a backbone node for routing. Note that in addition to the clusterhead, a cell can have other backbone nodes for routing. We can then employ a QoS routing algorithm to select an appropriate route consisting of cells. The route is "geographically fixed" even in the presence of moving MHs, and is referred to as virtual label-switched path (virtual LSP) since it is not composed of real network nodes or MHs. A signaling protocol is then employed to reserve the required resources from clusterheads on a cell-by-cell or region-by-region basis (instead of a node-by-node basis), and finally a QoS routing protocol routes the packets along the selected "geographically fixed route" with reserved resources (see Fig. 4). Since packets are always transmitted along a route with reserved resources (when these cells are not empty), QoS can be guaranteed in mobile ad hoc networks through GRACE. The control messages and mechanisms of ROC will then be utilized to find a feasible schedule between QoS MHs that may interfere with each other. More precisely, we essentially use ROC to reserve slots at future times on a slot-by-slot or group-by-group basis. More details concerning GRACE can be found in 17 , while details concerning the ROC-based reservation mechanism will be reported in the near future. GRACE can virtually guarantee QoS (with a high probability) when the ad hoc network is dense enough. However, in an ad hoc network that is not dense or a network where certain worst-case scenarios (such as those associated with the movements of MHs) may occur, we may employ cell protection or area protection (see Section 4) on top of GRACE to considerably reduce the probability for QoS guarantees to fail.

3.5 Moving Source or Destination QoS MHs In this subsection, we present the scenarios when an MH with QoS-guaranteed traffic, either a source MH or a destination MH moves too close other MHs with QoSguaranteed traffic. In such a scenario, GRACE along cannot guarantee QoS even in a dense ad hoc network. We may then have to reschedule the QoS slots, or rely on QoS adaptation as in cellular networks where an MH moves into a overutilized cell 16,17 . To avoid, or at least considerably reduce, the occurrence of QoS failures, we may employ another type of protection technique, called extension protection, to be introduced in the following section. Extension protection is useful in both ad hoc networks, wireless cellular networks, as well as several other wireless network architectures 15.

205 o o

o

o

o

%.

o o

o o o o

o o

o o

o

o

o

0

o

o

o

o

o

o

o

o o

o o

o

o

o o

o o

o

o o

o

o o o 0

o o^ o o 0

o o o o o o o

Figure 4: Routing along a geographically fixed route, where black circles represent source or destination mobile hosts, and white circles represent clusterheads that serve as backbone nodes for routing in this example. Radio resources are reserved by the QoS session through GRACE for the shaded cells traversed by the geographically fixed route. Note also that the clusterheads in the cells where the source or destination MH is located can be removed from the virtual LSP if the source/destination MH can transmit/receive to/from the clusterhead in the neighboring cell directly.

4

Ad Hoc Protection for Solving the Moving Terminal Problem

In this section we introduce several novel ad hoc protection techniques, including cell protection, geographical area protection, and extension protection. 4.1

Cell Protection with Virtual Reservation

Previous techniques for fault protection are mainly based on link protection, node protection, path protection, channel protection, or subpath protection. In order to guarantee continuous provisioning of QoS guarantees to existing connections with minimum disruption when a cell becomes empty, is intefered by noise, or has other types of link/node faults, we propose to employ cell protection, a new paradigm for fault protection and QoS guarantees in ad hoc networks. In cell protection, one or several protection label-switched subpaths (protection sub-LSP) with virtual reservation are established to bypass each cell that carries traffic with QoS guarantees. There are three major differences between our approach and previous protocols. The first, most distinguishable, difference is that we are protecting a cell instead of a link, node, or path, and we are not only protecting a cell against a malfunctioning node or link, but also an empty cell or a cell under considerable interference. To achieve this, each pair of neighboring cells of a protected cell are connected by one or several protection subLSPs, which will be activated when the protected cell can no longer provide the minimum required bandwidth to all the QoS sessions that traverse it (see Fig. 5 for two examples). The second difference is that we may use more than one path to protect a single subpath (with two links connecting to the protected cell), in contrast to a single protection path in previous protocols. This way the burden for accommodating

206 o

• \

o

> o

0

o

o

o

(a)

o

o

0

(c)

@-*—

-X

(b)

o

o

o

^

o o

~y

o

o

i

®—

o

o o

V

o (d)

Figure 5: Protecting QoS traffic traversing a cell that may become empty, faulty, or interfered. When the central cell of (a) becomes empty or faulty, the sub-LSP consisting of the three shaded cells of (a) is replaced by the sub-LSP consisting of the five shaded cells of (b). Similarly, when the central cell of (c) becomes empty, the sub-LSP consisting of the three shaded cells of (c) is replaced by the sub-LSP consisting of the three shaded cells of (d). If a cell only transmits to its four neighboring cells, we need at most 12 protection sub-LSPs to protect upto 12 sub-LSPs traversing the cell.

traffic traversing a protected cell that later on becomes empty, faulty, or interfered can be shared by multiple paths. More importantly, the risk that a protected cell and all its protection sub-LSPs fail at the same time (which may happen with a nonnegligible probability in mobile ad hoc networks when there is only a protection path) can be considerably reduced through multiple protection sub-LSPs. We refer to this capability as protection for multiple cell-faults. The third difference is that the virtual reservations of the protection sub-LSPs are utilizing the adaptable part 14 of the reserved bandwidth along the protection sub-LSPs, as illustrated in the following example (see Fig. 6). Consider an ad hoc network that is partitioned into rectangular cells. To protect the shaded subpath traversing the central cell in Fig. 6, we may establish two protection sub-LSPs connecting the two gray clusterheads. In this example, the bandwidth reserved by all the QoS sessions traversing the subpath is 60Kbps, including 24Kbps that is nonadaptable and 36Kbps that is adaptable under degradation. To protect the subpath against an empty or faulty central cell, we need to establish one or several protection sub-LSPs whose virtual reservations have an aggregate bandwidth equal to or larger than 24Kbps. We can first establish a protection sub-LSP through the three cells on the right column (i.e., traversing the shaded cells of Fig. 6). To do this, the current source clusterhead of the subpath to be protected (i.e., the upper gray circle) signals through the clusterheads located at the right column to the destination clusterhead (i.e., the lower gray circle). In this example, the link between the middle cell of row 1 (i.e., cell (1,2)) and its right neighboring cell has reserved bandwidth equal to 64Kbps, including 16Kbps that is nonadaptable and 48Kbps that is adaptable. The source Clusterhead (1,2) and Clusterhead (1,3) can then mark 16Kbps from the adaptable part of the

207 A 36Kbps NA 24Kbps

0

o 0

T f i

VR 16Kbps 48Kbps NA 16Kbps

o

VR 16Kbps A 0Kbps NA OKbbS

o VR 16Kbps A 8Kbps NA 8KbbS

0 VR 16Kbps A 32Kbps NA 32Khns

Figure 6: Establishing protection sub-LSPs with virtual reservations, where NA, A, and VR denote nonadaptable, adaptable, and virtual reservations, respectively. Note that two protection sub-LSPs are established to protect a single sub-LSP in this example. The sum of the virtual reservations of the two protection sub-LSPs has to be equal to or greater than 24Kbps, the nonadaptable part of the reserved bandwidth of the protected sub-LSP.

link reservation as the virtual reservation of this protection sub-LSP without requesting the downstream Clusterhead (1,3) for new bandwidth. We can then proceed to the link between the upper two cells of the right column (i.e., cells (1,3) and (2,3)). Since there was no reserved bandwidth over the link, Clusterhead (1,3) uses the ROC-based reservation mechanism to request for new resources from the resource broker of cell (2,3) (i.e., Clusterhead (2,3)). Assume that the request is successful, Clusterhead (2,3) should then record the schedule for the reserved 16Kbps. In the future if neighboring or nearby clusterheads request for bandwidth that may conflict with the reserved schedule, Clusterhead (2,3) has to reply with an OTS or object-to-reserve (OTR) message to reject such requests in order to keep the protection sub-LSP valid. We can then proceed to the next two links, where Clusterhead (2,3) has to request for additional 8Kbps for the required virtual reservation using the ROC-based reservation mechanism while Clusterhead (3,3) can simply mark 16Kbps from the adaptable part of the reservation of link ((3,3),(2,3)). Similar, we can establish another protection sub-LSP through the left three cells with virtual reservations at least equal to 8Kbps. When the protected cell becomes empty or when a fault occurs, the virtual reservations of its protection sub-LSP(s) are activated so that the sessions (of the proposed ad-hoc MPLS) with QoS guarantees can still be served with an acceptable, though degraded, quality. Since expensive radio resources do not need to stay idle unnecessarily for the protection purpose, while adaptive QoS can still be guaranteed once the associated virtual reservations are activated, our approach leads to effective protection without wasting expensive network resources. Note that when there are no appropriate paths with sufficient bandwidth to provide protection for all the QoS guaranteed traffic with separate virtual reservations, different protection sub-LSPs are also allowed to share a common virtual reservation for their protection. In such shared protection, a single cell-fault can still be tolerated without affecting the QoS guarantees, but multiple faults may cause some QoS

208

Figure 7: The dark area is protected by a sub-LSP (i.e., the path consisting of the gray cells). Any combination of faults or empty cells within the protected area can be tolerated by using the protection sub-LSP.

sessions to be disrupted in the worse case. Such sessions will then have to resort to restoration by signaling an alternative path. Similar strategy can be applied to path protection. More details concerning virtual reservations can be found in 17. 4.2

Geographical Area Protection and Extension Protection

We can generalize cell protection to protect an area, rather than a single cell, by using one or several protection sub-LSPs that bypass the area, leading to Geographical Area Protection (GAP), another new paradigm for protection in ad hoc networks (see Fig. 7). By combining GRACE and GAP with virtual reservation, QoS can continue to be guaranteed against any combination of faults within the protected area without wasting expensive resources. A third kind of ad hoc protection is to establish one or several sub-LSPs (with virtual reservations) that extend an existing LSP, by predicting possible locations of a source or destination MH (see Fig. 8). Note that when multiple locations are possible, multiple sub-LSPs with shared reservation can be established. More details will be reported in the near future. 5

Other Advantages and Extensions of ROC

In what follows, we present additional advantages that may be gained by adding the OTS mechanism to RTS/CTS. In the pure RTS/CTS scheme 9 , a transmitter relies on the CTS messages it received alone to determine whether its transmission will interfere with packets that are being received at the corresponding stations. There are two major problems that may occur in a protocol based on such a scheme alone: First, the CTS messages may be collided so that the intended transmitter is not aware of the existence of the corresponding transmissions that may be interfered. Second, other MHs that are receiving packets may move into the interference range of the intended transmitter, while the intended transmitter might not receive their CTS messages before since it was previously out-of-range. In , it was shown that the combined effect of

209 o o o o o o o o o % - ,-*o -»o-*? o o o o o o ! o o ° o 0 o o o o o o o 6 o o o p 0 0 0 o o o o o o °~~m" ^ T f c ^ 0 0 o o o o

o o o o o o o o

Figure 8: A virtual LSP with moving destination MH (i.e., the black circle) is protected by with a protection sub-LSP with virtual reservation. Note that the destination MH was predicted to move from the location corresponging to the gray circle to the black circle so that the virtual LSP was extended accordingly (by adding the two dark cells to the original virtual LSP).

the two aforementioned problems may considerably degrade the maximum achievable throughput of RTS/CTS-based protocols under certain assumptions. The proposed OTS mechanism essentially add an additional protection mechanism against the aforementioned problems, effectively mitigates their negative impacts. In 7 , a protocol based on busy tone n ' 1 2 was proposed to mitigate the aforementioned problems, and to completely solve the hidden terminal problem and the exposed terminal problem as RTS/CTS and ROC. However, previous busy-tone-based protocols reported in the literature 7 ' 11 ' 12 cannot provide QoS guarantees and cannot completely solve the moving terminal problem, though they may partially solve and mitigate the moving terminal problem defined in this paper. Busy-tone-based protocols cannot be applied to wireless networking environments where the maximum transmission radii of different mobile and base stations are different since busy tone cannot be forwarded or relayed; similarly, they are not suitable for wireless networks that allow variable transmission radii for the same device. Moreover, extra transceiver hardware is required to implement busy-tone-based protocols. However, we do believe that bust-tone-based protocols still derive further investigations due to their capability to continuously announce on-going transmissions, which may be useful in highly mobile networking environments with very small transmission radii. Therefore, incorporation of busy tone into ROC may deserve further investigations. We refer to the resultant scheme as ROC with bust tone (ROC/BT), which can also completely solve the moving terminal problem and has its respective advantages, but possibly at a higher implementation cost. Similarly, CSMA can also be combined with RTS/OTS/CTS dialogues, leading to ROC CSMA. More details will be reported in the near future. In 18 , we propose routing protocols that can take advantages of variable transmission radii to significantly increase network throughput and to reduce average power consumption. Another important advantage of ROC is that it can also support efficient transmissions with variable radii. We can augment ROC with an announce-to-send

210

(ATS) mechanism for an intended transmitter to inform all MHs that have received its RTS message whether it is going to transmit (as a follow-up message) and with a declare-to-send(DTS) mechanism for an intended transmitter to inform MHs within its actual interference or protection range so that the MHs that receive the DTS message can record the information or the associated schedule in its local database. This leads to a Request-to-send/Object-to-send/Announce-to-send/Declare-to-send(ROAD) protocol for support of various radii routing protocols 18. The RTS/OTS/CTS dialogues can also be used to negotiate the codes to be used in a CDMA system. For example, an intended transmitter can use an RTS message to inquire MHs within its interference or protection range whether the code(s) it selected are acceptable to them. If not, an objective MH can reply with an OTS message. If the intended receiver receives the code(s) and approves the usage, it can reply with a CTS message. Another application of ROC is to utilize the capability of ROC to reserve time slots or minislots for a cut-through switching technique, such as virtual cut-through, multihead cut-through13, or wormhole routing, to coordinate the transmissions of multiple flits or phits from nearby MHs in parallel. The success of such switching techniques may considerably reduce the latency for large-scale ad hoc networks. More details concerning ROAD and the applications of ROC to CDMA systems and cut-through switching techniques will be reported in the near future. 6

Conclusions

In this paper, we defined the moving terminal problem for high-mobility ad hoc networks with QoS-guarantee requirements. We proposed ROC to solve the moving terminal problem in mobile ad hoc networks, wireless LAN, PAN, or home networks. ROC can provision QoS guarantees in wireless mobile networks; it also naturally leads to variable-radius MAC protocols where the transmission radii can be dynamically changed to satisfy specific needs, such as increased network throughput and reduced average power consumption by using smaller transmission radii, or decreased latency for a time-critical session by using larger transmission radii. An implication of our results is that we demonstrate that QoS can be guaranteed based on distributed MAC protocols, even in high-mobility wireless networking environments. The implementation of ROC may be more cost effective than busy-tone-based protocols since no additional transceiver hardware is required. Due to the advantages of ROC and the importance of voice and multimedia applications in wireless networks, ROC may lead to an extension of the RTS/CTS-based CSMA/CA protocol of IEEE 802.11, or a new QoS-guaranteed MAC protocol for the Internet. References 1. Bertsekas, D.R and R. Gallager, Data Networks, (Prentice Hall, Englewood Cliffs, N.J., 1992). 2. Bharghavan, V., A. Demers, S. Shenker, and L. Zhang, "MACAW: A media access protocol for wireless LAN's," Proc. of ACM SIGCOMM'94, 212-225 (1994).

211

3. Corson, S., J. Macker, "Mobile ad hoc networking (MANET): routing protocol performance issues and evaluation considerations," RFC 2501 (1999). 4. Europe Telecommunications Standard (ETS), functional specification of the High PErformance Radio LAN (HIPERLAN) system (1995). 5. Federal Communications Commision, Part 15-RADIO FREQUENCY DEVICES, Subpart D-Unlicensed Personal Communications Service Devices, 47 CER0.1. 6. Garcia-Luna-Aceves, J.J. and C.L. Fullmer, "Floor acquisition multiple access (FAMA) in single-channel wireless networks," Mobile Networks and Applications, 4,157-174 (1999). 7. Haas, Z.J. and J. Deng, "Dual busy tone multiple access (DBTMA)-performance evaluation," Proc. IEEE Vehicular Technology Confi, 314-319(1999). 8. ISO/IEC 8802-11: 1999(E) ANSI/IEEE Std 802.11, Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications, 1 edition (1999). 9. Karn, P., "MACA - A new channel access method for packet radio," Proc. Computer Networking Confi, 134-140(1990). 10. Talucci, F, M. Gerla, and L. Fratta, "MACA/BI (MACA By Invitation): A receiver oriented access protocol for wireless multiple networks," in PIMRC'97 (1997). 11. Tanenbaum, A.S., Computer Networks, 3rdEdition, (Prentice Hall, N.J., 1996). 12. Tobagi, F.A. and L. Kleinrock, "Packet switching in radio channels. II. The hidden terminal problem in carrier sense multiple-access and the busy-tone solution," IEEE Trans. Communications, 23 1417-1433 (1975). 13. Yeh, C.-H., "A new switching technique for networks of workstations and LANs," Proc. Int'l Conf. Information Technology and Communication at the Dawn of the New Millennium, 439-448 (2000). 14. Yeh, C.-H., "Scalable QoS supports for multimedia applications in the nextgeneration Internet," Proc. IEEE Real Time Technologies and Applications Symp., 39-50 (2001). 15. Yeh, C.-H., "A versatile wireless network architecture with high throughput, low power, and QoS guarantees," Proc. IEEE Real-Time Technologies and Applications Symp., 114-116(2000). 16. Yeh, C.-H., "Medium access control with differentiated adaptation for QoS managementin wireless networks," Proc. IEEE Int'l Conf. Mobile and Wireless Communication Networks, 208-219 (2001). 17. Yeh, C.-H., "Scalable, adaptive, and reliable resource management in highspeed and mobile networks," Proc. IEEE Int'l Conf Computer Communications and Networks, (2001). 18. Yeh, C.-H., "Variable-radius routing protocols for high throughput, low power, and small latency in ad hoc wireless networks," Proc. IEEE Int'l Conf. Wireless LANs and Home Networks, (2001).

Part 8 Protocol Design and Mobility Support

VARIABLE-RADIUS ROUTING PROTOCOLS FOR HIGH THROUGHPUT, LOW POWER, AND SMALL LATENCY IN AD HOC WIRELESS NETWORKS Chi-Hsiang Yeh Dept. of Electrical and Computer Engineering Queen's University Kingston, Ontario, K7L 3N6, Canada E-mail: [email protected] In this paper, we propose several variable-radius routing protocols for achieving higher throughput, smaller latency at a given traffic load, and/or lower power consumption in ad hoc networks, by exploiting several unique characteristics of radio links. In particular, we propose small interference routing (SIR) that utilizes novel routing metrics and radius-oriented ad-hoc routing (ROAR) as a new paradigm for routing in ad hoc wireless networks. SIR and ROAR are generally applicable to existing routing protocols, such as on-demand, table-driven, and localized routing protocols, or new routing protocols, such as selective table-driven routing or embedded ad-hoc routing, to obtain their variants for achieving the aforementioned advantages. The resultant protocols are particularly useful for dense ad hoc networks, where the typical number of mobile hosts within the transmission range of a mobile host is not small.

1

Introduction

Recently, a new wave of interests in ad hoc networking1>3, where mobile hosts (MHs) are capable of communicating with each other without any pre-existing infrastructure, has emerged for wireless personal-area networks and home networks. The IEEE 802.11 standard 7 also includes ad hoc networks as one of its supported network architectures for wireless LANs. Moreover, wide-area ad hoc networks networks consisting of communication devices in automobiles, and ad hoc networks augmented with wired infrastructure have also been considered6'16'17. In an ad hoc network, a packet may require multiple hops to reach its destination due to the power constraints of portable devices as well as the requirements to achieve higher network throughput. Routing, therefore, becomes a research area of particular importance and interests for ad hoc wireless networks, which has received tremendous attentions from the industry and academia in the past few years 5 ' 8 ' 9 ' 10 ' 12 . A unique property of radio transmissions is that the power levels for transmitting from wireless devices can be adjusted. When higher power is used, the transmission radii will be increased so that the numbers of hops required to reach destinations can be reduced, decreasing the latency; on the other hand, when lower power is used, the transmission radii will be decreased so that the interference between MHs is reduced, increasing the network throughput. Several power control mechanisms have been widely employed in practice for energy conservation4'13'14, but controlling 215

216 transmission radii for improving routing efficiency, engineering network traffic and resource demands, or provisioning differentiated quality of service (QoS) has not been exploited thus far. In particular, many previous routing protocols aim at reducing the delay or the number of hops required for routing, which tends to use larger transmission radius and thus larger transmitting power, leading to lower network throughput in dense ad hoc networks, where where the physical distances between most neighboring MHs are small relative to the maximum transmission radii of those MHs. In this paper, we propose an evolutionary approach, Smallest Interference Routing (SIR), and a revolutionary approach, Radius-Oriented Ad-hoc Routing (ROAR), to achieve high throughput, small latency, and/or low power consumption for routing in ad hoc wireless networks. SIR utilizes new routing metrics to maximize network throughput and/or prolong battery lives by reducing interferences between routing tasks and the energy consumed, especially for areas or MHs where such resources are scarce. ROAR presents a new paradigm for routing, traffic and demand engineering, as well as differentiated service 2 in ad hoc wireless networks by adjusting the transmission radii appropriately. In particular, the resource requirements can be reduced at hot spots by reducing the interference ranges of transmissions at hot spots, on top of conventional techniques for traffic engineering such as detouring and flow control; on the other habd, within regions where the expensive radio resources are underutilized, we can increase the transmission radii so that the utilization of the radio channel is increased and the average number of hops required for routing is reduced, leading to smaller latency. We refer to such an extension to traffic engineering as demand engineering. Also, we can assign smaller transmission radii to time-noncritical sessions to achieve higher throughput for the network as a whole, while assign larger transmission radii to time-critical sessions to achieve smaller latency for those sessions, leading to a new tool for providing differentiated service quality for sessions with different QoS requirements, on top of conventional techniques based on priority and/or resource reservation15'18. ROAR is applicable to many existing or new routing protocols, including on-demand routing protocols 8,1 °, table-driven routing protocols 9 , localized routing protocols 12, Embedded Ad-hoc Routing (EAR) protocols 17, and Selective Table-driven Ad-hoc Routing (STAR) protocols (to be introduced later in this paper). Our work may lead to extensions to existing routing protocols, and may also lead to new routing protocols for ad hoc wireless networks. 2

Small Interference Routing (SIR) and wDa Routing

In this section, we propose SIR and several metrics for routing in ad hoc wireless networks with variable transmission radii. 2.1 Definition and Estimation of Ad Hoc Interference To define the ad-hoc interference for a transmission, a (unicast) routing task, or a group communication task in an ad hoc network, we first partition the physical space covered by the network into small squares, each called a tile (or, for 3-D space, into small

217

cuboids for 3-D space, each called a brick). Each tile (or brick) has a virtual price (e.g., $5 virtual dollars per usage plus $2 virtual dollars per 0.1 msec duration) for a user to pay if the user transmits a packet that interferes with that tile. Typically, there may be hot spots or downtown areas that have relatively higher virtual prices, and suburb areas that have relatively lower virtual prices or are even free; for simplicity, we may also set a uniform virtual price for all the tiles. The ad-hoc interference of a single transmission can then be defined as the virtual price it has to pay to all the tiles (or bricks) within its interference range. Since a (unicast) routiung task requires multiple transmissions for one or several packets, the ad-hoc interference of such a routing task is defined as the sum of the ad-hoc interferences of all the required transmissions. Similarly, the definition can be generalized to any communication task by counting the sum of the ad-hoc interferences of all the transmissions required by the task. The virtual price of a tile can be dynamically adjusted according to the current or expected traffic conditions or other criteria. For example, the tiles or bricks close to an access point may have heavier traffic loads, and thus higher virtual prices for transmissions that interfere with them; on the contrary, the tiles or bricks near the edge of the area covered by an ad hoc network may have lighter traffic loads, and thus lower virtual prices for transmissions that interfere with them. Note that we can assign similar virtual prices to tiles close to each other, but we may also make a tile free if there are not any MHs within it (or nearby). The virtual price of a tile can be a function of some parameters, such as the number of packets to be received or transmitted to/from the tile, the bandwidth reserved for MHs within the tile, the radio utilization of the tile, the weights and/or surcharges assigned by the transmitters, receivers, and/or other MHs within or near the tile, the strength of the signal when it propagates to that tile (e.g., when a CDMA-based MAC protocol is used), the starting time for the transmission (which may affect the uncertainty), the duration of the transmissions, the contract/program hold/enrolled by the user, etc. Several possible assignments of virtual prices will be provided in the following subsections to illustrate the ideas. Calculation of another special case of virtual prices or ad-hoc interferences, called ad-hoc utilizations, will be presented in Subsection 3.2. However, the best ways to assign virtual prices in order to achieve certain objectives, such as higher throughput, smaller latency, smaller protocol overheads, less expensive implementations, and/or longer battery life, require further investigations and experiments and are out of the scope of this paper. We believe that if the virtual prices appropriately reflect the traffic conditions of the tiles or bricks and the routing protocol in use selects routes with smaller ad-hoc interferences, then the network traffic can be engineered to avoid hot spots and reduce interference so that network throughput and service quality can be considerably improved as compared to previous protocols that are not adaptive to traffic conditions. 2.2 Small Interference Routing (SIR) In SIR, we select a route with the smallest ad-hoc interference using a shortest path algorithm such as Dijkstra or Bellman-Ford algorithms '. For MHs to obtain the in-

218

formation concerning ad-hoc interferences, we need functions for virtual prices that can lead to higher performance, mecahsnisms for the exchange of routing information and other parameters that determine the virtual prices, and the calculation of the adhoc interference for transmissions at a certain power level. In what follows, we use an example to illustrate the central ideas of SIR and to present a possible way of implementation. In this example, the virtual price of a tile is the total number of packets that are scheduled or expected to be received by the MH(s) located on the tile within a period of time. As a result, if a tile does not contain any MHs in it, the virtual price is $0, simplying the exchange and gathering of information as well as the calculation of ad-hoc interferences. We can use an extension of RTS/CTS protocols, where a CTS message contains the number of packets that are scheduled or expected to be received by the receiver. An active MH, especially one that serves as a backbone node, keep tracks of the ad-hoc interference incurred for transmissions with power level 1, 2, 3,..., and so on. We can then use the ad-hoc interference at power level j as the cost or length for a link if the link requires transmissions at power level j , and finally use a shortest path algorithm to find a small interference route. Since the ad-hoc interference is maintained as a function of the interference radius Dj, we refer to such an implementation as a special case of f(D) routing. Note that in addition to interference radii, we may also use the transmission radii as the parameter Dj in f(D) routing 17 . Note also that outdated or wrong information concerning ad-hoc interferences may cause inefficiency for routing, but not errors, and the estimations do not need to be very accurate for SIR to considerably improve the performance of previous protocols that do not minimize interferences between routing tasks. When CSMAl or its variants are used for the MAC protocol, the exposed terminal problem may occurs so that the transmissions from an MH may be forbiddened if the tile where it is located is interfered by another transmission. In such networking environments, the packets to be transmitted from the MH(s) on a tile should also increase its virtual price, since the performance is effected by such interference in reality. 2.3 wD2 Routing, wDa Routing, and Other Variants In an ad hoc network where MHs and traffic are (roughly) uniformly distributed within a reasonably small region, we can further simplify the calculation of ad-hoc interference. For example, an MH i can estimate the traffic load at its vicinity, and assign a weight Wj for its location. The ad-hoc interference for its transmissions with interference radius D; can then be estimated as wtD2. This leads to a simple implementation for SIR and is referred to as wD2 routing. The wieght w,- can be assigned according to other criteria such as the energy level or available processing resources. For example, an MH with a lower energy level may assign a larger weight w>,- to reduce its service as an intermediate node for other users, prolonging its battery life, while an MH with a higher energy level may assign a smaller or zero weight. This way, wD2 routing leads to routes that consume less

219

(A)

(B)

Figure 1: Interference or power requirements are proportional to the sum of the areas of the coverage ranges (i.e., the circles) for the corresponding transmissions. It can be seen that smaller transmission radius may lead to considerably smaller interference or energy consumption totally (i.e., from the source MH to the destination MH), even though the number of hops is increased, as shown in (B). power, and avoid MHs with lower energy levels, when we use transmission radius as the parameter D. Conventionally, traffic engineering is employed to avoid hot spots in order to increase network throughput and to improve service quality, while our approach based on wD2 routing demonstrates a new application of traffic engineering for ad hoc networks, where prolonging battery lives become the objective. When we assign w,- = 1 for all MHs i, a shortest path routing algorithm will naturally generate routing paths that require the smallest total energy consumption (assuming that the required transmission power is proportional to the square of the transmission radius). Figure 1 illustrates the amounts of energy required for routing between two MHs using different routing paradigms. Since the energy required for a transmission is proportional to the area of the circle corresponding to the coverage range, it can be seen that when they communicate directly (Fig. 1 A), the required energy is considerably larger than that when they communicate indirectly using smaller transmission radii (Fig. IB). Based on the metric w,D?, a QoS routing protocol can also lead to higher throughput and lower power consumption, after satisfying the constraints of the associated QoS sessions, such as the maximum tolerable delay. Note that the wD2 routing metric usually leads to routing paths that are very different from those derived by minimizing the hop counts, which constitute a routing metric commonly used in previous protocols in the literature. The rationale of our approach is that achieving higher throughput may be the most important goal for many applications of the future ad hoc networks with limited-bandwidth wireless links, and that when the achievable throughput is sufficient for the traffic load at a particular time, the power consumption will likely become the most important criteria, both of which are optimized when wD2 or its variants are used

220

as the routing metric. Although latency and jitter are critical for real-time applications, a QoS routing protocol can identify routing paths that satisfy the QoS requirements, and then select the one that minimizes the sum of WjDJ for all radio links i along the path. Therefore, there is no conflict between the proposed routing metrics and the requirements for QoS-aware applications. Finally, hybrid routing metrics including wD2 as part of their components may also be used, and may have their respective advantages. More details can be found in 17 by substituting D2 in 17 with wD2 from this paper. We can further generalize wD2 routing to wDa routing for other objectives or networking environments, such as different models for path loss. We may also substitute D2 in 17 with wDa from this paper. More details and simulation results will be reported in the near future. 2.4

Scalability of Dense Ad Hoc Networks

Most previous work in ad hoc networking consider it in the context of local area networks, while wide-area ad hoc networks with up to millions of nodes have also been considered 6 . In this subsection, we briefly discuss the scalability of ad hoc networks, taking into account the possibility of using variable-radius routing and MAC protocols 19

In a large and dense ad hoc network where MHs are distributed roughly uniformly, the maximum number of parallel paths is in the order of ®{\/N) when variable-radius routing and MAC protocols are used, where N is the total number of MHs. If fixedradius routing and/or MAC protocols are used, then the maximum number of parallel paths is upper bounded by 0(min(\/A/Z) max , VN)), where A is the physical area covered by the ad hoc network and Dmax is the maximum transmission radius of MHs in it. When the network is dense, \/A/Z) max is considerably smaller than \/N. Since the maximum throughput of an ad hoc network (or any other networks) is upper bounded by its bisection bandwidth, which is proportional to the maximum number of parallel paths when the traffic is uniformly distributed among MHs, an ad hoc network using variable-radius routing protocols such as SIR or ROAR (to be introduced in the following section) can achieve considerably higher throughput. Note that the aggregate network throughput of an ad hoc network grows rapidly with the number N of MHs in it (i.e., as a function of Q(y/N)), but not nearly as fast as the number N itself. So there will be a limit on the maximum number of MHs an ad hoc network can support, since the maximum throughput per MH decreases rapidly with the number of MHs (i.e., as a function of @(l/\/N)). Localized communication patterns may make ad hoc networks somewhat more scalable, but only to a limited extent. Thus, for ad hoc networks with a very large scale (such as tens of thousands or even millions), we conjecture that some form of infrastructure will be useful or even necessary. Also, many applications will require MHs to communicate with the outside world such as the Internet, constituting another reason for installing wired infrastructure for ad hoc networks. Several architectures that augment ad hoc networks with such infrastructures can be found in 16,17 and more details will be reported in the near

221

future. 3

Radius-Oriented Ad-hoc Routing (ROAR): A Generally Applicable Approach

In this section, we propose ROAR, a new paradigm for routing in ad hoc wireless networks, and apply it to a variety of existing or new ad-hoc routing protocols. 3.1 Demand Engineering: An Extension to Traffic Engineeringfor Wireless Networks Traffic engineering or traffic management is an important issue for networking, and has been intensively investigated for the Internet, multiprotocol label switching (MPLS) networks1 i, and asynchronous transfer mode (ATM) networks *. In this subsection, we propose demand engineering, an extension to traffic engineering for wireless networks with relaying capability such as ad hoc networks. In this paper, we refer to "demands" as the radio resources to be consumed, including the frequency bands in use, the required packet/session durations, and the interference areas resulting from transmissions. Note that only the former two types of resources are considered valuable in most previously work. For example, packet durations correspond to the consumed "time" of the radio channel as if they are "leases" of the expensive resources. In this paper, we follow a different networking paradigm where the transmission radii are variable and are selected to optimize latency and throughput, and we propose to treat interference areas (whose radii usually double the corresponding transmission radii) as another valuable resource, which corresponds to the consumed "space" of the radio channel (as "real estates" of the resources). This adds an additional dimension to the resource management of the radio channel, and can lead to considerably higher throughput, smaller latency, and/or lower power consumption. Demand engineering is a new concept for ad-hoc networking that can take advantages of the unique characteristics of radio transmissions. In addition to the standard techniques from traffic engineering such as detouring of traffic around hot spot, we show that demands for resources can be engineered in ad hoc networks by optimizing the transmission radii for all locations. More precisely, we may use smaller transmission radii where the radio resources are scarce (so that the interference ranges and thus the demands for resources become smaller), and larger transmission radii where the radio resources are abundant. In other words, not only the "location" of traffic can be engineered, but the "amount" of traffic can also be engineered in wireless networks. In the following subsections, we present several ways to estimate the demands for a region and mechanisms to engineered them appropriately. 3.2 Definition and Estimation of Ad-hoc Utilization The net ad-hoc utilization for a location is defined as the percentage of time it is interfered by transmissions; the net ad-hoc utilization for a tile or a region is defined as the percentage of time and area that the tile or the region is interfered by transmissions. For example, if 20% of the area of a region is interfered during time 0 and time 2, and

222

10% of its area is interfered during time 2 and time 5, then the net ad-hoc utilization of the region during time 0 to time 5 is given by 2 3 Un = 20% x - + 10% x - = 14%. The net ad-hoc utilization of a location, tile or region is always between 0 and 1. The gross ad-hoc utilization for a location, a tile, or a region is defined as the sum of the net ad-hoc utilizations resulting from each of the transmissions that interfere with it. For example, consider a scenario where there are n transmissions from MHs 1,2,..., n that interfere with a region during a period of time. If the interference range of transmission from MH i overlaps with a,% of the region (in terms of the overlapping area) for f,% of the time, then the gross ad-hoc utilization for the region is given by n

Ug =

^aiXti.

The gross ad-hoc utilization of a location, tile or region may be higher than 1. The sampled ad-hoc utilization for a location or an MH or device is the same as the net ad-hoc utilization of the location or the location where the MH or device is located; the sampled ad-hoc utilization for a tile or a region is defined as the weighted sum of the sampled ad-hoc utilizations of all the MHs and/or other measuring devices within the tile or region. For example, consider a region where there are n MHs 1,2,...,n within a region that have sampled ad-hoc utilizations U\,U2,—,Un, respectively, and weights wi,w2,...,w„, respectively. Then the sampled ad-hoc utilization for the region is given by n Us =

^WjXUi. 1=1

The sampled ad-hoc utilization of a location, tile or region is always between 0 and 1. Note that the weight w, may be equal to 0 so that the sampled ad-hoc utilization at some MHs can be ignored. We can also partition a region into irregular cells, and choose one MH from each cell to represent it. Then the weight w,- of an MH can be assigned to be equal to the area of the cell it represents. We may also associate the weights with other parameters such as the number of packets to be received or transmitted. Note that gross ad-hoc utilizations may be easier to calculate mathematically, while sampled ad-hoc utilizations and net ad-hoc utilizations may be easier to estimate when measurement is available. Note also that the preceding definitions can be generalized to CDMA or FDMA systems by multiplying the contribution from a transmission i with a weight w,, where w, may be the number of frequency bands or the number of CDMA sessions that can transmit at the same time.

223

3.3 Radius-Oriented Ad-hoc Routing (ROAR) In this subsection, we propose ROAR as a new paradigm for routing and traffic engineering in dense ad hoc networks. In ROAR, the desirable transmission radius (or interference range) at a location or region is dynamically controlled according to local traffic conditions and/or other criteria such as the energy level of a transmitting MH. For a heavily utilized region, the transmission radii should to be reduced (unless no appropriate paths can be found or the sessions have higher priority and/or require smaller latency) so that less resources are consumed and congestion and/or dropping of packets can be avoided or mitigated for the region; one the contrary, for an underutilized region, larger transmission radii are allowed to be used since we have abundant resources there and larger transmission radii may lead to smaller number of hops and thus smaller average latency. One way to regulate the transmission radii is to keep the net ad-hoc utilization bellow a certain threshold (e.g., 50% or 80%) for all regions, or the gross ad-hoc utilization bellow another larger threshold (e.g., 1 or 2). The rational behind such rules is that the average delay associated with transmissions in a region increases rapidly when its net or gross ad-hoc utilization approaches the corresponding maximum achievable utilization. Therefore, we should make sure that the net or gross ad-hoc utilizations for all regions are below a reasonable value most of the time so that none of the regions covered by the ad hoc network become congested or even saturated. As an example, consider two equal-size square regions R\ and R2 with m\ a n d m2 flows, respectively, that are traversing through them with average bandwidths b\ and b2, respectively. Then, in order for the gross ad-hoc utilizations of regions R\ and R2 to be roughly the same, the ratio of their desirable transmission radii becomes r\ : r2

RS

m^b-i :m\b\ = d2: d\

when the ad hoc network is dense, where d\ and d2 are the traffic density of regions R\ and R2, respectively. Note that the desirable transmission radius at a location may be different for different directions, which may be inversely proportional to the density of traffic in that direction. In addition to such initial estimations, dynamic control of the transmission radii may further improve the performance. For example, when the average queue length of a region becomes too large, we can reduce the desirable transmission radius so that more packets can be transmitted in parallel, reducing the number of packets accumulated within the region; otherwise, we may increase the desirable transmission radius to increase the utilization and to reduce the average hop count. If reducing transmission radii can no longer improve the throughput effectively, we should activate trafficengineering mechanisms to detour some traffic and/or flow-control mechanisms to reduce the injection rates of packets into the overutilized region. Other ways to regulate the transmission radii are also possible. For example, we can keep the sum of ad-hoc interferences (see Section 2.1) within a region during a

224

time unit below a reasonable threshold. But this may be more complicated to implement than the preceding two approaches. More details and experimental results, especially for the dynamic control of transmission radii, will be reported in the near future. 3.4 Applications of ROAR to Existing Protocols When the desirable transmission radius for a region is made available to all or most of the MHs within the region, existing routing protocols, such as table-driven, ondemand, or localized routing protocols, can be extended to incorporate ROAR. The resultant protocols are adaptive to traffic loads by appropriately selecting the transmission radii so that smaller latency can be achieved for any traffic load. Note that when the traffic is light, the desirable transmission radii can be as large as the maximum transmission radii of the MHs so that ROAR will degenerate to conventional routing protocols that minimizes hop counts; while when the traffic is very heavy, the desirable transmission radii are very small so that ROAR will behave in a manner similar to SIR, which minimizes interference between routing tasks. To obtain Table-driven ROAR (T-ROAR)protocols, we apply ROAR to corresponding table-driven routing protocols. This can be done by recording in the routing table routes that use links with physical distance approximately equal to, or upper bounded by, the desired transmission radius D, for the location i of the corresponding MH. Note that we can record multiple entries for a certain destination in the same routing table, some for smaller radii and some for larger radii. Therefore, when the traffic density changes so that the desired transmission radius D, changes, we use a different set of entries for routing without having to update the routing table right away. As a result, the overheads for T-ROAR protocols can be reduced and T-ROAR protocols can react to traffic changes promptly. Also, a session that requires smaller latency can choose to use the entries with larger radii to reduce the number of hops. This is acceptable as long as such sessions will not overload the corresponding region, while treating different sessions differently (e.g., by using different radii for them), according to their QoS requirements and/or the price they are willing to pay, may be important for some networking environments and is referred to as the differentiated routing discipline in this paper. To obtain On-demand ROAR (O-ROAR) protocols, we apply ROAR to corresponding on-demand routing protocols. One way to realize it is to use transmission radii somewhat larger than the desirable transmission radii, instead of the maximum transmission radius of MHs, to flood the control messages for route discovery. We can then select an appropriate route from the viable routes received by the destination. Note that this may considerably reduce the control message overheads, and is usually sufficient for routing under O-ROAR. However, a risk for this approach is that a better route may not be found since one or several of its links happen to be somewhat larger than the transmission radii used for flooding, and rerouting with larger transmission radii may be required to find a route with satisfactory performance, increasing the delay associated with route discovery. Another way to realize it is to use the maximum

225

transmission radius of MHs to flood the control messages for route discovery as in conventional on-demand routing protocols. We can then choose a route where most of the links have physical distances approximately equal to, or upper bounded by, the desired transmission radii of the associated locations or the corresponding MHs. Similarly, the differentiated routing discipline can be applied to O-ROAR for sessions with different priority and/or QoS requirements. To obtain Localized ROAR (L-ROAR) protocols, we apply ROAR to corresponding localized routing protocols. For example, to apply ROAR to depth first search (DFS) routing12, we first search an appropriate link that will move a packet to a position closest to the destination and has physical distance no more than /£>,-, a factor / (greater than 1) times the desired transmission radius D,. If we cannot find appropriate links, we should then relax the restrictions and try links that have larger physical distances by using larger / . Similarly, to support the differentiated routing discipline in R-ROAR, we simply use a larger / for sessions with higher priority and/or requiring smaller latency. The resultant protocol is called DFS-ROAR. Applications to various other localized routing protocols are also feasible. More details concerning T-ROAR, O-ROAR, L-ROAR, and other ROAR-based protocols based on geographical information will be reported in the near future. 3.5 Applications of ROAR to New Protocols With the availability of desirable transmission radii, we can also apply ROAR to new routing protocols, such as selective table-driven routing, embedded ad-hoc routing 17 , or SIR. Selective Table-driven Ad-hoc Routing (STAR) is a new hybrid routing paradigm where selected MHs, such as those that are (1) static or moving relatively slowly, (2) located at critical locations, and/or (3) satisfying other criteria such as sufficiently high energy level or favorable user options, maintain routing tables, while other MHs use other routing paradigms such as on-demand routing. The thresholds for selecting these table-driven MHs can be dynamically controlled according to traffic loads, network size, table size, and/or other criteria. To obtain Selective Table-driven ROAR (ST-ROAR) protocols, we apply ROAR to the table-driven MHs as in T-ROAR, and apply ROAR to another routing mechanism for the rest of MHs (such as on-demand routing), as we applied ROAR to other routing mechanisms (such as in O-ROAR). Embedded Ad-hoc Routing (EAR) is a another new hierarchical routing paradigm where backbone nodes are selected from MHs, and packets are mainly transmitted between backbone nodes for routing, except before a packet reaches a backbone node nearby and/or after it has been routed to a location near the destination and has to be relayed to the destination MH. A requirement for EAR is that the selected backbone nodes should be connected as, or subsume, a network topology with known connectivity, such as a triangular network17, a 2-D regular mesh, or a web or multiweb consisting of one or several stars or trees intersecting with concentric circles or polygons. To obtain Embedded-Route Radius-Oriented Routing (ERROR) protocols, we select MHs at

226

appropriate locations as backbone nodes so that the physical distances between neighboring backbone nodes are approximately equal to the desirable transmission radii for the region. Similar to T-ROAR and O-ROAR, we can have different transmission radii at a location for sessions with different priority levels and QoS requirements. On way to realize this in ERROR is to embed multiple topologies consisting of the backbone nodes of the ad hoc network. We can also combine ROAR and SIR to obtain Minimum-Interference-Route RadiusOriented Routing (MIRROR) protocols. In MIRROR, we only consider links that are within a reasonable range from the desirable transmission radii obtained in ROAR, and then select the routes with the smallest ad-hoc interference as in SIR. If the resultant routes or routing tables are not satisfactory, we can relax the restrictions and retry with a larger range. For example, to obtain routes with smaller latency, we may increase the upper limit of the range; to obtain routes with smaller ad-hoc interference, we may reduce the lower limit of the range. We can further generalize MIRROR by using other traffic engineering mechanisms to enhance or replace SIR, leading to Traffic-Engineered-Route Radius-Oriented Routing (TERROR). More details concerning STAR, ST-ROAR, EAR, ERROR, MIRROR, and TERROR will be reported in the near future. 4

Conclusions

In this paper, we proposed several variable-radius routing schemes for ad hoc wireless networks, including SIR, ROAR, and STAR. In particular, ROAR is a new paradigm for routing in ad hoc wireless networks, and leads to novel tools to traffic and demand engineering as well as differentiated service. We extended a variety of routing protocols based on ROAR, and obtain T-ROAR, O-ROAR, L-ROAR, ST-ROAR, ERROR, MIRROR, and TERROR protocols. The purpose of this paper is not to exhaust all possible routing protocols that utilize variable transmission radii or based on ROAR and SIR, but to initiate investigation into such a routing paradigm, which may have profound impacts on future ad hoc networking. We will refine the routing protocols proposed in this paper, and perform computer simulations to evaluate and compare their performance. We also plan to conduct experiments and measurements on the proposed protocols on a wireless network of workstations/laptops (wireless NOW) test bed. References 1. Bertsekas, D.R and R. Gallager, Data Networks, (Prentice Hall, Englewood Cliffs, N.J., 1992). 2. Blake, S., D. Black, M. Carlson, E. Davies, Z. Wang, and W. Weiss, "An Architecture for Differentiated Services," RFC 2475 (1998). 3. Corson, S., J. Macker, "Mobile ad hoc networking (MANET): routing protocol performance issues and evaluation considerations," RFC 2501, (1999). 4. Gomez, J., A.T. Campbell, M. Naghshineh, and C. Bisdikian, "Power-aware routing in wireless packet networks," Proc. IEEE Int'l Workshop on Mobile

227

Multimedia Communications, 380-383 (1999). 5. Haas, Z.J., M.R. Pearlman, and P. Samar, "The interzone routing protocol (IERP) for ad hoc networks," Internet Draft draft-ietf-manet-zone-ierp-01.txt (2001). 6. Hubaux, J.-R, J.-Y. Le Boudec, S. Giordano, et al, "Towards mobile ad-hoc WANs: Terminodes," Proc. IEEE Wireless Communications and Networks Conf., (2000). 7. ISO/IEC 8802-11: 1999(E) ANSI/IEEE Std 802.11, Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications, 1 edition, 1999. 8. Johnson, D.B., D.A. Maltz, Y.-C. Hu, and J.G. Jetcheva, "The dynamic source routing protocol for mobile ad hoc networks," Internet Draft draft-ietf-manetdsr-05.txt, (2001). 9. Perkins, C.E., and P. Bhagwat, "Highly dynamic destination-sequenced distance-vector (DSDV) for mobile computers," ACM SIGCOMM, 234-244 (1994). 10. Perkins, C.E., E.M. Royer, and S.R. Das, "Ad hoc on-demand distance vector (AODV) routing," Internet Draft draft-ietf-manet-aodv-08.txt, (2001). 11. Rosen, E.C., A. Viswanathan, and R. Callon "Multiprotocol label switching architecture," RFC 3031, (2001). 12. Stojmenovic, I., M. Russell, and B. Vukojevic, "Depth first search and location based localized routing and QoS routing in wireless networks," Proc. Int'l Conf. Parallel Processing, 173-180 (2000). 13. Stojmenovic, I. and X. Lin, "Power-aware localized routing in wireless networks," Proc. Int'l Parallel and Distributed Processing Symp., 371-376 (2000). 14. Toh, C.-K., "Maximum battery life routing to support ubiquitous mobile computing in wireless ad hoc networks," IEEE Communications Magazine, 39,138147 (2001). 15. Yeh, C.-H. and E.A. Varvarigos, "The scalable networking scheme for highspeed networks," Proc. IEEE Int'l Conf. Communications (ICC'2000), 13351342(2000). 16. Yeh, C.-H., "A versatile wireless network architecture with high throughput, low power, and QoS guarantees," Proc. IEEE Real-Time Technologies and Applications Symp., 114-116 (2000). 17. Yeh, C.-H., "A new network architecture for mobile communications," Proc. World Multi-Conference on Systemics, Cybernetics and Informatics, 367-374 (2001). 18. Yeh, C.-H., "Scalable, adaptive, and reliable resource management in highspeed and mobile networks," Proc. IEEE Int'l Conf. Computer Communications and Networks, (2001). 19. Yeh, C.-H., "ROC: A wireless MAC protocol for solving the moving terminal problem," Proc. IEEE Int'l Conf. Wireless LANs and Home Networks, (2001).

AN OVERVIEW AND COMPARATIVE EVALUATION OF WIRELESS PROTOCOLS ANTOINE MERCIER Ecole Centrale d'Electronique, 53 rue de Grenelle, 75007 Paris, France E-mail: [email protected] PASCALE MINET INRIA, Domaine de Voluceau, 78153 he Chesnay, France E-mail: [email protected] LAURENT GEORGE, GILLES MERCIER LllA, University Paris XII, 122 rue Paul Armangot, 94400 Vitry-sur-Seine, France E-mail:{george, mercierj @univ-parisl2.fr Wireless networks come in force into the market of computer communications. No doubt that in short future, those technologies will be developed in smart engineering systems. Unfortunately, several standards coexist (e.g. IEEE 802.11, Bluetooth, HomeRF, HTPERLAN), leading to various types of products. The purpose of this paper is to present an overview of those standards and to establish a comparative evaluation. We mainly focus on the Medium Access Control protocols. Those protocols generally support two modes of operation, a random access mode for asynchronous traffic and a polling mode for real-time synchronous traffic. This article proposes an analysis of the ability of each protocol to meet user Quality of Service requirements.

1

Introduction

Emerging need of wireless communications drives the development of new standards and products that bring drastic changes to the wireless communications landscape. New and emerging standards such as IEEE 802.11, Bluetooth, HomeRF, HTPERLAN type 1 and 2, allow wireless communications between mobile and desktop computers, cellular telephones, and other hand-held digital devices. Customers are faced today to a wide variety of wireless technologies, systems and vendors. We present a brief description of these five wireless protocols to help end users or future customers in their choice. To establish comparative criteria, we consider the traffic submitted by an application to a wireless network. Each traffic type requires a certain level of Quality of Service (QoS) affecting the quality of transmission perceived by the user. The QoS can be examined from two perspectives: user and network perspective. From user perspective, QoS requirements are based on user's perception of the level of service delivered. The user QoS is translated into a network QoS according to a mapping function. The network role is to ensure that the desired QoS is met by means of management and control functions. In this paper, we first focus on Medium Access Control (MAC) protocols for the five wireless studied

228

229 standards, particularly on mechanisms provided to support QoS requirements according to the user point-of-view. We describe how user level QoS requirements are mapped to low level network. We then study the adequacy of these five wireless standards w.r.t. different application constraints. Finally, we conclude and give some open issues for future work. 2 2.1

Wireless standards and protocols Generalities

Before presenting protocols, we describe different network topologies and medium access type used in Wireless Local Area Networks (WLAN). WLAN can be built with or without an infrastructure leading to either an infrastructured topology or an ad-hoc topology. In an infrastructured topology, there are specific devices called Access Point (AP). An AP not only controls the wireless communication of the devices in its radio range, it also provides access to a wired network. Moreover as the radio coverage is limited and generally does not allow to reach any device in the WLAN, routing solutions should be implemented in the WLAN. In an infrastructured topology, the wired network is used for routing between devices depending on different APs. In an ad-hoc topology, routing can be done at the MAC level. We can distinguish two medium access types: centralized and decentralized. In a centralized medium access type, a central entity shares the medium between the devices in its radio range, by means of a polling scheme. It ensures that all transmissions between devices under its control are contention free. Generally, this access type is used for synchronous traffic. Moreover, with this access type, it is easier to implement power saving mode, as the polling scheme can account for the sleeping period of battery operated devices. In a decentralized medium access type, each device is independent and competes for the medium access. Random access schemes are used. In a wireless environment, collision detection is not possible; that is why a collision avoidance scheme is used. Examples are given by Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) and Elimination Yield-Non Pre-emptive priority Multiple Access (EY-NPMA). An infrastructured topology uses a centralized medium access, where the control entity is the AP. An ad-hoc infrastructure generally uses a decentralized medium access; however it can also use a centralized one, where a central entity is elected. We now study the PHY and MAC layers of five wireless standards that are IEEE 802.11, Bluetooth, HomeRF, HIPERLAN type 1 and 2. 2.2

IEEE802.il

The IEEE 802.11 [6] protocol defines an ad-hoc network architecture for devices within mutual communication range of each other. When devices wanting to

230 communicate belong to different coverage area, the distribution services are needed. They are provided by Access Points. Access points are needed for bridging with wired networks and forming a WLAN. If an Access Point exists, a centralized medium access scheme can be used; otherwise a decentralized medium access is used. IEEE 802.11 MAC can provide to users two service types for data transmission: (i) asynchronous with contention period, provided by the Distributed Coordination Function (DCF) and (ii) synchronous with contention-free period (CFP) provided by the Point Coordination Function (PCF). The integration of the two access methods is obtained with a superframe including two successive periods. O P Repetition Interval Contention Free Period

— •

Contention Period

PCF

DO-

' — Foreshorl CF Period PCF

Variable Length (per Superframe) Figure 1. Basic MAC frame structure IEEE 802.11

Different Inter-Frame Spaces (IFS) have been introduced. These inter-frame spaces of different duration are a way to implement prioritized access. Some authors have studied different mechanisms to implement it within the superframe [14]. The DCF implements the basic access method CSMA/CA with random back off counter that provides fair access to all stations without the need for centralized control. The back off counter is bounded by two configurable parameters: the Short Retry Limit (SRL) and the Long Retry Limit (LRL) which let the control of the latency and the reliability of pending messages [1]. To deal with the hidden node problem, the 802.11 MAC protocol includes an optional mechanism based on the exchange of short control frames: ^Request To Send (RTS) frame sent by a potential transmitter to the receiver and a Clear To Send (CTS) frame sent from the receiver in response to the received RTS frame. The effectiveness of the RTS/CTS mechanism depends upon the length of the packet to be protected [15]. A RTSjhreshold fixed by the user could be applied to disable the mechanism for shortest frames. The PCF, which implements a polling access method, is available only in an infrastructured topology where an AP is used. The PCF relies on the service provided by the DCF, because devices requesting for synchronous communication have to use DCF rules to notify the AP. The AP periodically polls devices giving them the opportunity to transmit. The maximum duration of CFP (CFP_Max_Duration) and its repetition rate (CFP_Rate), defined as a number of superframe, should be determined according to the characteristics of traffic that has to be conveyed by the AP. Other polling based MAC mechanisms, more efficient than Round-Robin, can be implemented to support real-time communications (see for instance [10] [13]).

231

2.3 Bluetooth Bluetooth devices [2] use fast frequency hopping for data transmissions. Any two Bluetooth devices within range of each other can set up an ad-hoc connection, so establishing a small network called piconet. Each piconet consists of a unique master which selects a frequency hopping sequence for the piconet and controls the access to the medium. Other participants of the piconet, called slave units, are synchronized to the hopping sequence of the piconet master. Up to seven active slaves units may form a piconet. Multiple overlapping piconets can co-exist because each piconet uses a different hopping sequence. Piconets can be interconnected via bridges to form a bigger network called scatternet. However, to support forwarding data between devices from different piconets, routing mechanism must be implemented at upper layers [12]. To communicate, Bluetooth MAC implements a Time Division Duplex (TDD) scheme, splitting the channel into slots alternately used for transmission and reception between the piconet master and its slaves. The slots for any master-to-slave communication are immediately followed by slots for slave-to-master communication. A slave is only allowed to transmit in a given slot if the master has addressed it in a preceding slot. Moreover any direct communication between slaves is impossible. Piconet master has to manage a polling scheme between slaves according to RoundRobin scheme or more efficient scheme as studied in [3]. master

slave 1

master

1*625 us, 3*625 us or 5*625 us master master

!

slave 2

i

Hap

1Hop

| Hop SCO packets

|

1

Hop

Bs#'' W A w ..;Af^1M&.Vi

! Hop -

! slave s x

, t#P :,1

example of Hop ">uW-sl« packet in ACL communication

Hop

Figure 2. TDD and timing in Bluetooth communication

Two types of service for data communications are provided to the users: (i) asynchronous, called Asynchronous Connectionless (ACL) and (ii) synchronous named Synchronous Connection Oriented (SCO). ACL service is used for the transmission of data burst. The master unit manages entirely the ACL connection bandwidth: a master polls a slave which may answer during following slot(s). The master defines the maximum packet length granted to the slave; this length is based on criterion such as the quantity of data to be transmitted or the number of successive slots available in the presence of SCO traffic. A maximum poll interval can be negotiated as the maximum time between subsequent transmissions between the master and a particular slave. It is used to support bandwidth allocation and latency control. The SCO service supports time constrained point-to-point connections. The length of a SCO packet cannot exceed one time slot.

232

A best-effort QoS is generally applied to communications. A user can request a QoS Guarantee for a traffic defined by its flow spec parameters such as e.g. Token Rate, Peak Bandwidth... All these options have to be negotiated during connection establishment. 2.4 HomeRF Based on the Shared Wireless Access Protocol (SWAP) 1.0 [11], HomeRF architecture is a combination of (i) a managed network for synchronous and centralized services and (ii) an ad-hoc network that provides asynchronous service. A unique AP, generally a gateway to a wired network, manages centralized services, synchronizes and controls communications on the medium according to MAC rules. If the Access Point is not available, the devices may create an ad-hoc network where control of the network is distributed between all the devices. For communication services, a user may choose between two services types: synchronous and asynchronous. The SWAP MAC includes a Time Division Multiple Access (TDMA) scheme to support synchronous traffic, as well as a CSMA/CA scheme derived from DCF in IEEE 802.11. This last scheme supports asynchronous traffic and control frames (e.g. request for synchronous traffic). The MAC protocol uses a superframe incorporating two contention-free periods and one contention period. The access mechanism used during each contention free period is TDMA, while CSMA/CA is used during the contention period. Hop

Contention Free Period for

« - * •

•*

B

D4

03

U3

U4

Contention Free Period in normal transmission

retransmission

*•

Hop

**

Contention period CSMA/CA access mechanism

»••*-*•

D4

D3 U4

Dl

D2 U3

V2

U!

Beacon tramc Dl, 1)2,1)3: Access Point to station 1,2,3 communication slot 111, U2, U3: Station 1,2,3 to Access Point communication slot Figure 3. Basic MAC frame structure in HomeRF

Each contention-free period is divided into 6 pairs of fixed length slots. The first slot of each pair is used to transmit data from the AP to a device. Respectively, the second is used for communications from a device to the AP. The second CFP, at the end of the superframe, is used for data transmission, while the first CFP at the start of the superframe is used for optional retransmission of any data not received or incorrectly received in the previous frame. To adapt the bandwidm to its requirements, a user can set die repetition rate of its allowed slots under CFP mechanism.

233

2.5 HIPERLAN/1 The MPERLAN/1 Standard [4] works independently from any fixed infrastructure as an ad-hoc network but allows access to conventional wired networks. HIPERLAN/1 supports the ad-hoc topology with a multi-hop routing capability at the MAC layer. With this capacity, packets are forwarded from a source to a destination which cannot be reached directly. This routing protocol [8] is a proactive one: it maintains a route for each destination in the wireless network. It belongs to the link state family, in which the broadcast of topology information has been optimized. The Multipoint Relay set, subset of the one-hop devices, allowing to reach all two-hop devices, has been introduced for that purpose: only the Multipoint relays of the device having broadcast a packet forward this packet, making this routing protocol efficient. The MAC layer of HIPERLAN/1 uses a form of priority in granting access to the medium. The access scheme used is called EY-NPMA. It operates similarly to CSMA. Basically, when the channel is assessed to be clear, packets ready for transmission are submitted to a prioritization phase during which the packets with the highest priority are selected. The two criteria used for establishing the priority levels of a packet are user level priority (as described in the traffic constraints (see section 3.1)) and residual packet lifetime (normalized to the number of hops to reach the final destination) deduced from the packet deadline given by the user. The mapping function of the couple (user level priority, normalized residual lifetime) into the medium access priority is given in [9]. 2.6 HIPERLAN/2 The HiperLAN/2 standard [5] is mainly defined for infrastructure based topology where an Access Point manages the wireless network. The AP controls devices in its range with connection-oriented links. This centralized topology is combined with an ad-hoc capability: unlike some centralized protocols where two devices attached to the same AP cannot communicate directly, HIPERLAN/2 has an option allowing a direct communication between such in range devices. Several logical links can be established between devices and a priority, mapped from the user level priority, is attached to each link. This priority can be translated into traffic classes according to IEEE 802.lp [7].When a device wants to transmit data, it must first send a resource request to the AP which grants resources in the next frames according to the traffic, the others requests and its own transmission buffers. A request contains the number of packets to be sent, the emission frequency, the link quality requested. These parameters are determined according to the application's needs.

234 Downlink slots

Uplink slots

Retransmission information slot Direct link slots Random access Control slots (optional) slot Figure 4. Superframe structure in HTPERLAN/2

The MAC protocol in the HIPERLAN/2 standard is based on TDMA principle. The basic frame structure has a fixed duration and incorporates several communications slots. The AP can adapt the frame composition to traffic variation, to adjust access delay. The HTPERLAN/2 standard provides a set of functionalities to support Quality of Services requirements such as dynamic frequency and data rate selection, link adaptation, handover, error control mode. 3 3.1

Adequacy between studied wireless protocols and application constraints Comparative criteria

We now focus on any application run on a wireless network. This application has constraints w.r.t. the network supporting it. We can distinguish diree types of constraints: global ones, MAC related and traffic related. Among the global constraints for a wireless network, we have: •

• • • • • • •

nomadicity of devices: some or all devices can change their location and want to keep the same virtual environment (e.g. services offered to the user); the mapping of the virtual environment to the real one depends on the physical location of the device. Such a device must be joined as soon as it has reached its destination whatever its destination. mobility of devices: some or all devices can move while mey are communicating. Service continuity must be provided. power management: battery operated devices must be able to save power when they are inactive. number of interconnected devices, global throughput. confidentiality: a minimum requirement is given by the WEP (Wired Equivalent Privacy), where a wired LAN equivalent data confidentiality is provided to users of wireless networks. easy installation and maintaining. cost effectiveness: networking cost must be low for three reasons: first in case of low cost device, second, in case of a high number of interconnected devices, third when the application lifetime is short. ability to interconnect with other networks (wired or not).

235

The constraints expressed by the application with regard to a wireless MAC concern: • • • • •

the need of a wired infrastructured topology. centralized/decentralized medium access. centralized communication: any device always communicates with a specific device or not. In the latter case, a wireless network ensuring a direct communication between devices will be preferred. communication with out of range devices: a routing protocol will be needed if out of range devices must communicate. the hidden node problem.

An application generates on the network different types of traffics and expresses traffic oriented constraints that are: • •

coexistence of different types of traffics: an application can be characterized from a network point of view as a set of weighted elementary traffics. The weight of each traffic depends on the application. dynamics of the traffic distribution: a wireless environment is subject to frequent changes. Consequently the way the traffic is distributed over the devices changes in space and in time: for instance devices previously in communication can become inactive. In such environments, the latency to adapt to atimeor space variation of the traffic distribution is an important parameter to consider.

The traffics produced by an application can be decomposed into elementary traffics. Each elementary traffic can be characterized by: • • • • • •

3.2

throughput, message size. user level priority. time constraints: they can be expressed by the maximum end-to-end delay affordable, the maximum tolerated jitter, the maximumtimeto establish or release a connection. dependability: it defines the network behaviour in case of failure (message or connection loss, device failure...). type of communication: unicast, multicast, broadcast. message arrival law: periodic (with constant inter-arrival time), sporadic (with inter arrival time greater than or equal to the minimum inter-arrival time), aperiodic (with only one arrival), bursted (with the duration and the period of the burst). Recapitulative table

Table 1 shows how the studied wireless standards account for the different application constraints: global, MAC related and traffic related.

236 Table 1. Application constraints applied to studied wireless networks IEEE 802.11 Nomadicity

Global constraints

Mobility

Power management Number of interconnected devices Global throughput Confidentiality

MAC constraints

Easy installation

IIomeRF

Bluetooth

HIPERLAN1

HIPERLAN2

Partially achieved by theAP Master mobility affects piconet configuration Must be managed by upper layer

Achieved by the MAC routing protocol

Partially achieved by the AP

Yes (for slow, medium speed)

Yes, if roaming by the fixed LAN

With awake patterns

With awake patterns and buffering devices

With awake patterns and AP buffers messages during sleeping period

7 slaves per master

-

256

1.6 Mbps

1 Mbps

23.5 Mbps

Authentication -(-encryption Immediate for ad-hoc

Key Sharing + encryption

Encryption

Yes

Immediate for ad-hoc

Key sharing + encryption Immediate for adhoc Yes

Partially achieved by the AP

Only within the AP coverage

Yes, if managed by upper layers

No

With awake patterns and AP buffers messages during sleeping period

With awake patterns and AP buffers messages during sleeping period

127

127

1,2,11 or 54 Mbps Authentication + WEP Immediate for ad-hoc

54 Mbps

Easy maintaining Cost effectiveness Ability to interconnect with other networks

Yes

Yes

Yes

Yes

Medium 200$ Wim bridge for IEEE802 LAN, with router otherwise

Low 5 to 20$ With PPP protocols or with bridge for IEEE802 LAN

High

High

With bridge for IEEE802 LAN, with router otherwise

With internal functionalities to support packet and cell based traffics

Wireless topology

Ad-hoc or infrastructured

Infrastructured

Ad-hoc

Ad-hoc or infrastructured

Decentralized for async. traffic Centralized otherwise Decentralized for asynchronous traffic only

Low-medium 10 to 100$ With bridge for IEEE802 LAN, with router otherwise Ad-hoc or infrastructured with only one AP Decentralized for async. traffic Centralized otherwise Decentralized for asynchronous traffic only

Centralized

Decentralized

Centralized

Centralized

Decentralized

Centralized communication

No

No

Yes

No

Communication wirti out of range devices Solution for hidden node problem

Yes if upper layer routing

Yes if upper layer routing

Yes if upper layer routing

Yes

Yes if upper layer routing

RTS and CTS frames

Only one AP used

Frequency Hopping

Active signalling

Each neighbour uses a different channel frequency

Medium access type for infrastructured topology Medium access type for ad-hoc topology

Centralized with election of an AP for the control of the medium No if optional direct communication

237

Unicast Multicast Broadcast: between in range devices

Unicast Multicast Broadcast: between in range devices

Unicast: between a master and a slave Broadcast: master only

Unicast Multicast Broadcast

Multicast Broadcast: between an AP and devices Unicast: between any two devices in range of the AP

End-to-end delay guarantee

Between two in range devices with polling

Between two in range devices with polling

In a piconet only, with polling

Deadline oriented but probabilistic access

Between two in range devices with polling

Bounded jitter

Between two in range devices with polling

Between two in range devices with polling

Probabilistic

Between two in range devices with polling

Loss control

With ARQ for unicast packets

With ARQ for unicast packets

Between two in range devices with polling With ARQ for unicast packets

No

No

No

With ARQ for unicast packets Translated into a MAC access priority

With ARQ for unicast packets Translated into a rank of connection establishment

Dynamics of traffic distribution

No delay for async. traffic change Update of the polling table for sync, traffic change

No delay for async. traffic change Update of the polling table for sync, traffic change

Update of the polling table for any traffic change

No latency

Update of the polling table for any traffic

Coexistence of different traffic types

With a superframe: polling for sync, traffic, contention for asynchronous

With a superframe: polling for sync, traffic, contention for asynchronous

Polling for sync, and async. traffics Jitter for async. traffic function of maximum polling interval

Contention for both synchronous and asynchronous traffics

With a superframe: Polling for any traffic

Traffic constraints

Unicast Multicast Broadcast

3.3

User level priority

Adequacy analysis

We summarize here the main results obtained from Table 1. If the application is such that all communicating devices are in range, any studied standard is possible. However, if the communication is not centralized (i.e. in the application any two devices can communicate directly), the Bluetooth solution and the HIPERLAN/2 solution without the direct communication option are less efficient. Indeed with these two standards, any communication is not only controlled by the AP or the master but it is received by this specific point and then resent to its destination. On the other hand, if the application is such that out of range devices have to communicate and there is no infrastructure, routing is needed between all the devices of the WLAN. Routing can be implemented either at the MAC layer (e.g. mPERLAN/1) or at the upper layer (e.g. routing in MANET working group at IETF). If an infrastructure exists, each AP is in charge of routing messages whose sender or destination does not belong to the devices it manages. This routing is achieved by a layer higher than two. In the same way, mobility can be handled at layer 2 (e.g. HIPERLAN/1) or at layer 3 (e.g. Mobile IP). As bandwidth is a killing resource in wireless networks, the

238

routing traffic is reduced by limiting as much as possible the number of devices propagating the routing information. Generally, the traffic induced by an application has a synchronous component and an asynchronous one. HIPERLAN/1 is the only wireless standard handling these two components in a uniform way: the bandwidth is shared according to the message priority accounting for the message deadline and the user level priority. All the other standards have resort to two different MAC access schemes: polling for the synchronous traffic and contention for the asynchronous one. When die synchronous traffic is well known and is not subject to frequent changes, a polling scheme can guarantee low jitter and transmission delay for this traffic. However, it must be carefully studied, in order to maximize bandwidth utiUzation: the slots allocated to polled devices having nothing to transmit must not be wasted. On the other hand, a MAC access scheme based on contention adapts itself immediately to traffic changes. However the contention scheme must be improved to manage different priorities, enabling the message with the highest priority in the collision to be transmitted. The coexistence between both access schemes is made possible by the following rule: the synchronous traffic is served first and the asynchronous traffic is served only if there is enough bandwidth left by the synchronous traffic. This rule is acceptable by an application, only if this application has no real-time constraint attached to asynchronous traffic. Indeed, this traffic can suffer long queuing delays increasing the jitter and the end-to-end transmission delay. 3.4

Open issues

At the end of this study, we can point out some open issues. These issues can be grouped in six topics: •

•

•

MAC access scheme and QoS requirements: Hidden node: How to find an efficient solution to this problem inherent to wireless network? Mapping of user priority and real-time constraints: The real-time constraints can be attached to synchronous traffic and to asynchronous one. In all studied standards but HfPERLAN/1, asynchronous traffic is transmitted only if the synchronous traffic as left enough bandwidth. How can it be improved? How to differentiate asynchronous traffic with different real-time constraints? How to decrease the latency to establish a new synchronous traffic? Availability: How to tolerate the failure of an Access Point? End-to-end QoS in multi-hop wireless networks: How to account for the distance a message has to experiment to reach the final destination? Can it be mapped into a medium access priority? How to extend the QoS properties obtained locally to multi-hop? How to keep determinism when an AP to a wired network is used? Routing in multi-hop wireless networks:

239

• •

•

4

Efficiency of routing done at the MAC level vs. at the IP level: From the strict performance point-of-view, implementing routing at layer 2 is a better solution. However, it supposes that routing algorithms are implemented in the network controller firmware. This is a more expensive solution than a layer 3 solution generally implemented in software. QoS routing in multi-hop wireless networks: This is needed in order to ensure an end-to-end QoS in such networks. QoS routing must select a route that offers the features (e.g. bandwidth, delay) required by the requested QoS. Mobility in multi-hop wireless networks: How to guarantee that the QoS is ensured while the device is moving? Mobility and real-time constraints are difficult to conciliate. Design approaches in wireless networks: In a hierarchical approach, more intelligence is put in the AP, less complexity in the other devices. As the number of APs is small with regard to the number of other devices, such an approach allows to reduce the networking cost of the application. In a fully decentralized approach, all devices have the same intelligence. In order to reduce the amount of control traffic, they can elect some among them to achieve particular. Compatibility of different solutions: A problem to be addressed concerns the compatibility of the different products. It can be discussed at two levels: (i) Coexistence and interference of different wireless networks in the same environment; (ii) Interconnection of different wireless networks: a compatible MAC frame format would make it easier. Conclusion and perspectives

Wireless networks are now easy to deploy and the available bandwidth becomes acceptable for most applications. Moreover, with wireless networks, mobility has become a reality. Such networks encounter an important success and are of great interest for smart applications. In the same time, the offer for wireless products is increasing, based on different wireless protocols. Moreover, the standardization is quickly evolving (e.g. HomeRF version 2.0, Bluetooth discussed in the IEEE 802.15 Personal Area Network Working Group). This paper proposes a comparison of those wireless protocols. We have focused on the first two layers of wireless standards that are IEEE 802.11, HomeRF, Bluetooth, HTPERLAN/1 and HTPERLAN/2. Our motivation is to help a designer in the choice of a wireless standard. Some issues are still open. They concern QoS management, routing, mobility, design approach and compatibility. For instance, all studied protocols except HTPERLAN/1 favour synchronous traffics versus asynchronous ones, which are transmitted only if there is sufficient bandwidth left by synchronous traffics.

1. Aad I. and Castelluccia C, "Introducing service differentiation into IEEE 802.11", in Proceedings of the Fifth IEEE Symposium on Computers and Communications, ISCC2000 (Antibes, France, 2000). 2. Bluetooth Special Interest Group, "Specifications of the Bluetooth system 1.0b, Volume 1 : Core", http://www.bluetooth.com (1999). 3. Das A., Ghose A., Razdan A., Saran H., and Shorey R., "Enhancing Performance of Asynchronous Data Traffic over the Bluetooth Wireless Ad-hoc Network", IEEE INFOCOM (Anchorage, USA, 2001). 4. ETSI, "ETSI STS-RES10 Technical Committee. Radio Equipment and Systems: High Performance Radio Local Area Network Type 1, Functional specifications" (1996). 5. ETSI, "ETSI TS-101 761-1 Technical Specification Broadband Radio Access Networks (BRAN), HIPERLAN Type 2. Data Link Control (DLC) Layer Part 1: Basic Data Transport Functions" (2000). 6. IEEE 802.11, "LAN MAN Standards Committee of the IEEE Computer Society. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications" (1997). 7. IEEE 802.1D, "IEEE Standard for Local Area Network MAC Bridges" (1998). 8. Jacquet P., Minet P., Miihlethaler P. and Rivierre N., "Increasing Reliability in CableFreeRadio LANs, Low Level Forwarding in HIPERLAN", Wireless Personal Communications (1996), Vol. 4, pp. 51-63, Kluwer Academic Publishers. 9. Jacquet P., Minet P., Miihlethaler P. and Rivierre N., "Data transfer for HIPERLAN", Wireless Personal Communications (1996), Vol. 4, pp. 65-80, Kluwer Academic Publishers. 10. Kopsel A., Ebert JP. and Wolisz A., "A performance Comparison of Point and Distributed Coordination Function of an IEEE 802.11 WLAN in the Presence of RealTime Requirements.", in Proceedings of 7th Intl, Workshop on Mobile Multimedia Communications, (Japan, 2000). 11. Negus Z. J., Stephens A. and Lansford J., "HomeRF and SWAP: Wireless Networking for the Connected Home", ACM SIGMOBILE Mobile Computing and Communications Review (1998), Vol. 2, pp. 28-37. 12. Salonidis T., Bhagwat P., Tassiulas L., and LaMaire R., "Distributed Topology Construction of Bluetooth Personal Area Networks", IEEE INFOCOM (Anchorage, USA, 2001). 13. Veeraraghavan M., Cocker N. and Moors T., "Support of voice services in IEEE 802.11 wireless LANs", IEEE INFOCOM (Anchorage, USA, 2001). 14. Wang Y. and Bensaou B., "Priority Based Multiple Access for Service Differentiation in Wireless Ad-hoc Networks", Networking 2000 (Paris, France, 2000). 15. Weinmiller J., Woesner H., Ebert JP. and Wolisz A., "Analyzing the RTS/CTS Mechanism in the DFWMAC Media Access Protocol for Wireless LANs", IFIP TC6 Workshop Personal Wireless Communication (Prague, Czech Rep., 1995).

SUPPORTING MOBILITY IN DISTRIBUTED PROXY SERVER ARCHITECTURE HYUKJOON LEE Dept. of Computer Engineering, Kwangwoon University, Seoul, Korea E-mail: [email protected] KWANGSUE CHUNG Dept. of Electronic

Communication Engineering, Kwangwoon University, Seoul, E-mail: [email protected]

Korea

KISUPKIM LG Electronics, Inc, Seoul, Korea E-mail: [email protected]

One of the most widely accepted approaches in improving user experience with WWW over wireless mobile networks has been transcoding proxy strategy. However, this approach does not only suffer from a serious drawback of being a potential point of failure but also introduces a new bottleneck in the network. A straightforward solution for this problem is to distribute multiple proxy servers over the entire service area and to have each proxy server serve clients within its designated service area only. To support mobility of the clients that move from a region served by one proxy server to another region served by another proxy server, a new handoff message protocol was designed and implemented. Using this proxy server handoff protocol the client on a mobile host remains connected to the proxy service. The performance of the proposed approach is shown to scale up very well as the number of proxy servers increases with very small amount of overhead in the handoff processing.

1

Introduction

Recent advances in cellular communication technology have enabled a new kind of Internet service called "Wireless Internet." However its quality of service is considered to be much lower than what users experience with the most of wired counterparts. Therefore, research and development have been continuously driven towards overcoming problems related to the poor characteristics of wireless media such as narrow bandwidth, high bit error rate (BER), limited display size, short battery life, and so on, [1,2,3,4]. Transcoding proxy servers have been widely accepted as a practical means of reducing delay in wireless Web access. The main functions of these proxy servers are caching and transcoding, also known as distillation. Transcoding may help users experience accelerated end-to-end network speed through a lossy compression This work was supported in part by the University Research Program of the Ministry of Information & Communication of the Republic of Korea.

241

242

approach which reduces the amount of transferred data through the wireless link at the expense of the lower quality of images in resolution, size, the number of color levels, and so on. Caching also helps reduce the transfer delay by shortening the traveling distance of HTTP request and reply messages. In BARWAN project Fox, et al. developed a proxy-based transcoding system called Transend that performs real-time wireless adaptation for resource-poor wireless mobile hosts by reducing the size, color level, resolution, and so on, of the WWW objects such as JPEG and GIF images [5]. In [6], Housel et al. presented a similar work. They used a unique architecture that consists of a client, an interceptor, and a server. Jung, et al. tried to optimize the HTTP procedure for a transcoding proxy [7]. One of the most significant drawbacks of proxy servers is that it may become a single point of failure or a network bottleneck. One straightforward approach to solve this problem is to use multiple proxy servers that can share the load of service requests. This approach, however, cannot solve the problem completely since the link between the proxy servers and gateways or interworking functions may still be congested. An alternative approach is to take advantage of the inherent distributed processing characteristics of cellular networks, that is, to distribute the proxy servers over the entire service coverage area. In this approach each proxy server may be made to provide service for the mobile hosts (MH) within the assigned subarea. As a result, both the workloads of the proxy servers and the network traffic can be distributed. Another advantage of this approach is that frequently requested data in location dependent services can be locally cached in the nearby proxy servers. Each sub-area can be the coverage area of one or more base stations (BS) depending on the system configuration of a wireless mobile data service network. Here, if the MH moves between the cells served by two different proxy servers a handoff between the proxy servers must be carried out. Therefore a proxy handoff message protocol must be defined. This paper presents our work in the design and implementation of a distributed proxy server system and a proxy server handoff message protocol. There are a number of previous research works that inspired our work. In [8], Gwerzman, et al. proposed to deploy distributed cache servers to decentralize the network traffic from a web server. When the number of user accesses to a particular file exceeds a specific threshold, the web server selects a cache server that uses the least network bandwidth and replicates the file to the chosen cache server. Generally, the cache server nearest to clients is chosen for replication. In [9], Haahr, et al. introduced mobility gateway with the ability to support CORBA-based application. While a mobile host moves to a new mobility gateway and connects to it, the handoff between old mobility gateway and new mobility gateway occurs. During handoff, the loss of data was prevented using the handoff message protocol. The rest of this paper is organized as follows. In section 2, the proxy server handoff protocol is described. In section 3, we explain the design of the distributed proxy server and the client. Then, in section 4, we present the result of

243

implementation and performance evaluation. Finally, we draw conclusions and discuss future work in section 5. 2

Proxy server handoff protocol

If the MH handoffs between two base stations served by two different proxy servers, the TCP connection from the MH to the first proxy server must be terminated and a new TCP connection must be established between the MH and the second proxy server. The HTTP session that is terminated due to the connection break must be resumed with the newly connected proxy server. We call this process the proxy server handoff. The proxy server handoff process begins with the mobile host collecting its location information through the interface to the bearer of the MH. We assume this location information contains a base station ID and the new proxy server can be determined from this base station ID. We call the proxy server connected to the MH before the handoff the old proxy, and the proxy server connected to the MH after the handoff the new proxy. A proxy server handoff in the middle of a HTTP session may cause some of the objects to be blocked from reaching the MH. For seamless web service, the terminated session must be restored and the blocked objects must be transmitted to the MH. The proxy server handoff message protocol is designed to handle the proxy server handoff. 2.1

Proxy server handoff processing

Proxy server handoff processing consists of three steps: (i) proxy handoff initiation, (ii) proxy data synchronization, and (iii) continued download from the web server. In the first step proxy server handoff, the client agent of the MH establishes a new TCP connection to the new proxy server. It then issues a handoff request message to the new proxy server. In the second step the new proxy server sets up the list of files that are remaining to be received by the MH. Some of these files may have been received, transcoded and cached by the old proxy server but have not been transferred to the MH yet. The new proxy server then requests the old proxy to transfer these files. This step helps reduce the delay associated with fetching the objects from the origin server. In the final step, the new proxy server also sends HTTP requests for the rest of the files belong in the same page to the origin server. These files are transcoded and stored in the cache as well. There are three different cases of proxy server handoff processing depending on whether the old proxy has received some part of a page as shown in figure 1.

244

2.2

Proxy server handoff message format

There are three types of proxy handoff protocol messages: (i) proxy server handoff request message used by the client, (ii) proxy data synchronization request message, and (iii) proxy data synchronization reply message. Figure 2 illustrates the format of the proxy HO request message. Total size represents the total length of the handoff

Query on BS ID BSID Handoff Req. HTTP Req. HTTP Res. (HTML) HTTP Res. (HTML) HTTP Req. t

HTTP Res.

Data

(a) VLR/HLR

Mobile Host

New Proxy

^ Query on B S I D

Handoff Req.

HTTP R is. (HTML) Proxy Data Sync. R q. Msg. ProxyDala Sync. R< sly Msg.

(b) VLR/HLR

Mobile Host

New Proxy

Query on BS ID

HTTP R S. (HTML) Proxy Data Sync. Ri q. Msg. Proxy Data Sync. R( ply Msg.

(c)

Figure 1. Proxy server handoff processing procedures: (a) when the MH has received no files of a web page, (b) when the old proxy has received a part of a page, and (c) when the old proxy received the entire page

245 Total Size (2 byte)

New Proxy IP Address (4 byte) User Request

List of Files Received by MH

Offset (4 byte)

Figure 2. Proxy handoff request message

Total Size (2 byte)

No. of Files (1 byte)

File Name 1

Expiration Time 1 (8 byte)

File Name n

Expiration Time n (8 byte)

Figure 3. Distilled data request message Total Size (4byte)

File Name

File Size (4byte)

Data Figure 4. Distilled data message

request message and old proxy IP address specifies the IP address of the old proxy. User request specifies the HTTP request message sent before proxy server handoff. This information is used to restart the HTTP session. List offiles received by MH specifies the objects received by the MH before the proxy handoff. Using this information the proxy server can extract the URLs of all objects within a page and determine the files that have not been received by the MH. In case the connection between the old proxy server and the MH is broken in the middle of transferring an object, the transfer must be resumed for the rest of the object. Information on the amount of the object received prior to the disconnection is provided in received bytes. In the proxy data synchronization request message (figure 3) No. of files (up to 255) specifies the number of files to be transferred from the old proxy to the new proxy. File name contains the names of files that have not been received by the MH. Expiration time is the time used in cache invalidation required by HTTP. The third type of message is the proxy data synchronization reply message (figure 4). File Name is the name of a file that the old proxy server transmits to the new proxy server. Size of Data specifies the size of the file and Data contains the data itself.

246

3

Distributed proxy server system

Both the client and proxy server system must be extended to handle the proxy handoff processing procedures. Hence, modules for handling the proxy handoff messages should be added to the usual HTTP proxy server and the client on MH, respectively. Figure 5 shows the overall system architecture. 3.1

Distributed proxy server

The distributed proxy server is an extension of a HTTP proxy server [10]. Thus, it is built on top of a standard web proxy server. As figure 5 illustrates, the main components of the distributed proxy server are the HTTP handler, the request message handler, the proxy data synchronization handler, the transcoded data dispatcher, the transcoder, the cache manager, and the file list manager. The message handler receives either HTTP messages or handoff request messages from the MH and branches to the HTTP handling routines or the handoff processing routines accordingly. The file list manager generates the list of files that are pending to be received by the MH and passed onto the proxy data synchronization handler or the HTTP handler. The proxy data synchronization handler sends request for the files remaining in the old proxy server. The HTTP handler makes HTTP requests to the origin web server. The transcoded data dispatcher transmits the transcoded data to the MH. 3.2

Client agent

The handoff processing module within client agent consists of the location information manager, the proxy server connection manager, the handoff request message generator, and the file handler. We assume that the location information Proxy Server

Web Server

Other Proxy Server

HTTP Handler

' Request *, Message L-— Handler J

M

Proxy Data Sync. Handler

File List Manager

Client

Transcoder

Client Agent v

Cache Manager

ranscoded" Data Dispatcher^

Figure 5. Architecture of the distributed proxy server

j

247

manager be provided with the information on the base stations from the bearer. Based on this information the location information manager can determine whether proxy server handoff is needed. In case of proxy server handoff, the IP address of the new proxy server is determined and passed to the proxy server connection manager. The proxy server connection manager then establishes a TCP connection to the new proxy server. The handoff request message generator generates the proxy server handoff request message and sends it to the new proxy server. The file handler passes the files received from the new proxy server to the client browser. 4 4.1

Implementation and performance Implementation

The proxy server was implemented based on a traditional HTTP proxy server in GNU C on a UNIX machine. The client agent was implemented as a local proxy in C++ on a Microsoft Windows machine. 4.2

Performance

Since we did not have a real cellular system for our performance test, we configured our testbed based on a wired network as shown in figure 6. We used four proxy servers to compare the performance of the distributed system of two, three, and four proxy servers against a single proxy server. Also, a load-balancing server was used to compare the performance of the distributed proxy servers against a cluster of multiple proxy servers with TCP packet redirection. In this case, the proxy handoff processing mechanisms were turned off. On each of eight PC's, up to four clients were run at the same time so that a total of up to 32 clients were run. Each client was configured to emulate handoffs of up to the number of proxy servers minus one. We emulated handoffs of MH's using a handoff emulation module to generate

P'°xy 4

Host 2

Host 4

Host 6

Host 8

Figure 6. Network configuration for the performance test

4

8

16

32

Number of hosts

Figure 7. Performance test results comparing the single proxy server, the distributed proxy servers, and the clustered multiple proxy servers with load-balancing

event signals for the clients. We measured the total retrieval time, while increasing the number of proxy servers from one to four, the number clients from 2 to 32, and the number of handoffs from zero to three for each of several HTML pages. This was repeated 20 times to determine the average retrieval time for each case. We present here only one test result due to space limitation. The test results for other pages were consistent. The size of the test page was 50.5 Kbytes and contains four image objects of 49.0 Kbytes in total. The test results in figure 7 indicate that a considerable amount of improvement in performance was achieved with the distributed proxy servers over that of a single proxy server. It also shows that the distributed proxy servers slightly outperform the load-balanced multiple proxy servers. We expect that more significant improvement could be made if the test was performed in a realistic setting where the link to the load-balancer becomes a bottleneck. It is clear from the figures that the performance of distributed proxy servers improves as the number of MH's increases. We expect this improvement rate may gradually saturate as the number of proxy servers approach a certain number. Although we have not come up with a single analytical model for the optimal number of proxy servers, we are investigating it. Also, the amount of overhead caused by the proxy handoffs can be said to be quite small

249

Number of hosts Figure 8. The average retrieval time of the distributed proxy servers with or without the proxy . handoff processing

because the amount of increase in the average retrieval time as the number of proxy handoffs increases was small. In figure 8, we compare the average retrieval time of distributed proxy servers with or without the proxy handoff processing to study the tradeoffs between the overhead in the proxy handoff processing and the overhead in the retransmission of blocked objects from the origin server. Based on this graph, we can claim that the utilizing the distributed caches with the proxy handoff processing and transcoding can improve the performance. 5

Conclusion

To solve the problem of being a single point of failure or a bottleneck with traditional proxy server approach, we proposed to use multiple proxy servers that are distributed over a wide area. We designed and implemented not only the distributed proxy server and client system, but also the proxy server handoff protocol that enables handoff between the proxy servers. We tried to minimize the amount of file transfer during the proxy server handoff processing by making the old proxy server send its transcoded and cached objects to the new proxy server.

250

We evaluated the distributed proxy server system and the proxy server handoff protocol that we implemented on a wired local area network setting. The test results show that we can achieve significant amount of improvement when we adopt the distributed proxy servers and the proxy server handoff protocol while keeping the amount of overhead fairly low. References 1. P. Sinha, N. Venkitaraman, R. Sivakumar, and V. Bharghavan, "WTCP:A Reliable Transport Protocol for Wireless Wide Area Networks", in Proceeding of the fifth Annual ACM/IEEE Internl. Conf. on Mobile Computing and Networking, Seattle, Washington, USA, 1999. 2. WAP forum, "WAP White paper", http://www.wapforum. org. 3. I-mode, http://www.nttdocomo.com/imode/. 4. cHTML, "Compact HTML for Small Information Appliances", http://www.w3.org/TR/1998/NOTE-compactHTML-19980209. 5. A. Fox, S. D. Gribble, E. A. Brewer, and E. Amir, "Adapting to network and client variability via on-demand dynamic distillation," Operating Sys. Rev., vol. 30, no. 5, Oct. 1996. 6. B. Housel and D. Lindquist, "Webexpress: A system for optimizing Web browsing in a wireless environment," in Proceeding of the Second ACM/IEEE International Conference on Mobile Computing and Networking, 1996. 7. K. Ham, S. Jung, S. Yang, H. Lee, K. Chung, "Wireless-adaptation of WWW Content over CDMA", Sixth IEEE International Workshop on Mobile Multimedia Communications (MOMUC99), San Diego, Nov. 1999. 8. J. Gwerzman, M. Seltzer, Harvard University, "The Case for Geographical Push Caching". 9. M. Haahr, R. Cunnimgham and V. Cahill, "Supporting CORBA Applications in a Mobile Environment", in Proceeding of the fifth Annual ACM/IEEE Internl. Conf. on Mobile Computing and Networking, Seattle, 1999. 10. HTTP (Hypertext Transfer Protocol), http://www.w3.org-/Protocols/.

Part 9 Interoperability and Co-existence

IAPP ENHANCEMENT PROTOCOL* JIN XIAOHUI, LI JIANDONG [email protected],

[email protected]

(Phone number: 0086-029-8201002,

0086-029-8201337)

(ISN National Key Lab., Information Science Institute, Xidian Univ., XVan, 710071, China) Abstract: In this paper, 3 kinds of the DS topologies are introduced. Then 3 IAPP protocols available are analyzed and compared from their complexity and robustness. A new IAPP enhancement protocol based on DSC is proposed, and the handover process and query process are described and be proved has better robustness than those protocols before. The transfer agent is introduced to solve the problem of roaming among IP sub-networks.

1

Introduction

IEEE 802.11 standard describes the implementation of Distribution System (DS) "outside the scope of this standard" [1], which makes the implementation of Wireless LAN (WLAN) flexible. However, this flexibility comes at a cost - WLAN products from different vendors cannot interoperate, especially among Access Points (APs) [2]. The role of the DS is to facilitate the delivery of frames between wireless stations and other stations in the network. A DS is typically formed by a collection of APs connected by an Ethernet II network or IEEE 802 layer-2 network. While APs operate independently, the reliability and capability of the DS is greater if they cooperate using an Inter Access Point Protocol (IAPP).

supported by 863-371project, Nature Science Fund of China (69872028)

253

254

2

Three typical DS Topologies and three IAPP protocols

Since IEEE 802.11 standard was adopted, a large number of manufacturers have developed and marketed APs. The majority of these products are constructed as bridges, with the logical portal function contained within each AP. The logical DS function is implemented over a standard IEEE802.3 LAN. IEEE802.3 LAN has many kinds of topologies, which divided into 3 main kinds, as shown in Figure 1-3. _ 802.x LAN Portal!

Portal2

802.x LAN

Figure 1. One collision domain and one broadcast domain domain

Figure 2. Several collision domains and one broadcast

Figure 3. Several collision domains and several broadcast domains Figure 1 is the simplest topology, which only contains one collision domain and one broadcast domain. In Figure 2, the network is divided into several collision domains by Bridges/Switches, layer 2 devices. Because broadcast frames can be transferred transparently by bridge/switch, the network is still considered as one broadcast domain. Figure 3 introduces Routers, the layer 3 devices, which normally refuse to

255 forward broadcast frames. The whole network is composed of many broadcast domains. [3] has described the simple handover scene shown in Figure 4, which can be used into the circumstance as shown in Figure 1. In Figure 4(a), (1) when a Mobile Station (MS) moves into a new AP's BSA, it sends a reassociation request to the new AP and informs the old AP, which it associated before. (2) The new AP sends back a reassociation response. (3) the new AP sends unicast/multicast handover request to the old AP. (4) me old AP sends or does not send a disassociation to the MS in its BSA depends on different implementation. This algorithm is very simple. But its drawbacks is that the potential loss of handover request and without the confirmation of handover response may results in simultaneous multiple associations of MS. So, a handover response is introduced into Figure 4(b). Before the new AP sends back a reassociation response to the MS, it must wait for the handover response from the old AP. The multiple association is vanished, but an additional loss probability of handover response and an excess delay to MS's reassociation are introduced. Usually, the old AP is announced by MS in its reassociation.request. But in many cases it's maybe unreliable. In Figure 4(a) and Figure 4(b), the validity of the old AP cannot be proved. To solve the unreliability of old AP, a regional database is introduced into Figure 4(c). The new AP gets some necessary information from the database not the old AP. By introducing the database, the multiple association and the invalid old AP are not exist. The complexity of Figure 4(c) is same as Figure 4(b), a little complex than that of Figure 4(a). The handover processes of [3] are simple and not robust, and only suits for the simple topology, 1 collision domain and 1 broadcast domain. 1

regional .rtataJiasft

*ers£ New AP

v,

Old AP

New AP

Old AP

MS

MS

(a)basic handover process

(b)handover with response from old AP

New AP

Old AP

U

MS

(c)handover with response from regional database

Figure 4. The simplest handover protocol

256

The algorithm of [3] is so simple because its implementation based on a very simple DS. All APs are directly connected, i.e., every AP can hear others. But in most situation DS's implementations are more complex, like Figure 2 and Figure 3 show. Sometimes there are some layer-2 devices, such as Bridges/Switches, divide the DS into several broadcast domains. And sometimes it must be considered when a MS roams from an IP subnet to a new IP subnet. The consistency of DS and MSs' smooth roaming are all must be considered. [4] proposes the concept of DSC (Distribution System Coordinator), which configures and maintains DS by sending DsBeacon frame. DSC can be selected from all DSCC (DSC Candidate) according to their priority. A DSCC is a normal AP. [4] describes how to set DSCC priority, select DSC, handover with DsNotify, and inquire with Dslnquiry & DsResponse. But all above processes don't make best advantage of DSC. The DSC, which only used to configure DS dose not participate in the handover and inquiring processes at all. And [4] also does not describe how to transit old DSC to new DSC smoothly with the DS continuity. In addition, although [4] presents the handover and data exchange processes in routed-broadcast and routed-non-broadcast distribution conditions, the processes are so complex that cannot really solve the question when a MS roams from an IP sub-network to another IP sub-network. The proposal of [4] can handle very complex DS, such as APs' add/remove, APs' authentication, etc. and the whole system is more robust. Of course, its implementation is more complex than that of [3], and need to define some new frame format, such as Dslnquiry and DsResponse.

1

fp'

1*—-—_

^-~- >\N \

Switch Bridge j

New AP (a) handover process

I Old AP

1

(b) inquiry process

Figure 5. The role of master AP Because of the useful of database that had shown in [3], the concept of master AP is introduced in [5], which is shown in Figure 5. The role of master AP is very like the

257

regional database in figure 4(c) and also regarded as a fixed DSC. In Figure 5(a), (1) after receives a reassociation request from a MS, the new AP sends a unicast handover to the master AP. (2) the master AP sends back the relative information of that MS. (3) then the new AP sends a new unicast handover request to the old AP. (4) the old AP sends back a handover response to confirm the handover of that MS. Then the new AP can send the reassociation response out. But sometimes it's risk to set a fixed master AP, especially in bad circumstance. When the master AP fails, the whole WLAN cannot work further because all the handover process and inquiry process cannot finish. In Figure 5(a), before the new AP send reassociation response to MS, it will spend longer time for new AP to wait information about MS's old AP from the master AP and MS's previous association from its old AP. The query process is shown in Figure 5(b). (1) when the old AP wants to know the information of a MS, it can sends unicast query to the master AP. (2) the master AP sends back the address of the new AP which that MS now associated. (3) the old AP send unicast query to the new AP. (4) the new AP replies all the necessary information. This procedure can be to query MSs or when the old AP detects that a MS had disassociation without handover request and disassociation request. The complexity of [5] is between [4] and [3]. Its simple is at the cost of the robustness of the system. 3

A new IAPP enhancement protocol

Select DSC from all DSCC has great flexibility. And in a DSC work period, the central control of handover and inquiry processes by DSC is very useful. Based on these understanding, a new IAPP enhancement protocol is proposed in this paper. 3.1 the selection and maintenance of DSC We can select DSC by two methods. The first method likes [4] describes: select DSC according to all APs priority, which value varies from 0-3. An AP with priority of zero will never become DSC. The AP with the highest number priority will become DSC. If two APs have the same DSC priority, then the AP with higher MAC address will become DSC. The second method is: according the different topology of DS and the purpose of traffic balance, the administrator can specify the

258

sequence of all APs becomes DSC and store the sequence into all APs. Only the former DSC fails, the latter can attempts to become DSC. The DsBeacon frame sent by DSC should contain a DSCC field, which contains all active DSCCs (APs) according their priority sequence, i.e. DSCCl, DSCC2...DSCCn. The destination address of DsBeacon is IAPP multicast address and the source address is the MAC address of the DSC. Because of the limited number of APs, the DSCC field will not increases DsBeacon length largely. After receiving the DsBeacon, DSCCl will sends back an ACK to indicate its activity and can replace the DSC. The others DSCC will not respond. If the DSC receives the ACK from DSCCl, the DSC will keep the DSCC field in the next DsBeacon, else after 3 DsBeacon intervals the DSC can assume the DSCCl fails, replace DSCCl to DSCCn with DSCC2 to DSCCn-1, and then send a new DsBeacon with new active DSCC sequence out. Now the new DSCCl, i.e. the former DSCC2, should reply an ACK for the DsBeacon. So, we are sure that the DSCCl in DSCC field is always active and can be DSC quickly. If the DSC fails, all DSCC cannot receive DsBeacon in consecutive 3 DsBeacon intervals. The DSCCl will try to become DSC and sends DsBeacon immediately. If all DSCCs don't receive DsBeacon in consecutive 3*n DsBeacon intervals, the DSCCn, which the last DsBeacon indicates, will become DSC and send DsBeacon immediately. By this method, we can assure the administrator's configuration and management right and the DSC continuity when the former DSC fails. If a new AP powers on, it sends DsStart.request to declares to establish a DS for a particular SSID or receives DsBeacon from the DS with the SSID appointed. If there already exists a DSC in the DS, the AP should receive a DsBeacon from it. After comparing its priority with the existing DSC's, if the AP's priority is lower, the AP will become a normal AP in the DS by sending a DsJoint request to the DSC; else the AP will send a Dslnquiry to the DSC and get all information about the WLAN, such as the DSCC sequences and all MSs locations, instead of transferring to a new DSC immediately as [4] shows. When receiving the DsResponse from the DSC, the AP sends back an ACK and becomes the new DSC. After receives the ACK, the previous DSC degrades to a normal AP and stop

259 sending DsBeacon. By DsInquiry/DsResponse exchange, the new DSC can get all information necessary and assures the continuity of the whole WLAN. 3.2 handover process and inquiry process For the central controlled mode, the DSC not only can record all the association/reassociation information, but also can play a more important role in handover and inquiry processes.

(a) handover process

(b) inquiry process

Figure 6. The handover process and inquiry process of IAPP enhancement protocol • In Figure 6(a), (1) after receives a reassociation request from MS, the new AP sends handover request to the DSC. (2) The DSC searches its database, replies a handover response contains the information about the reassociation MS, such as the old AP and the QoS parameters, and accedes the new AP send reassociation response to the MS. (3) after send out handover response, the DSC sends DsNotify to the old AP to informs it of the MS's new place. The destination address of DsNotify is the old AP and the source address is the MS. Using MS as the source address can update all the involved bridges/switches. (4) After receiving the DsNotify, the old AP deletes the MS entry invalid locally, replies the DSC an ACK frame, and forwards all buffered frames destined to the MS to the new AP. This method results in some delay for the old AP to delete the invalid entry. It is fortunate that the delay is very Utile and that all the frames destined to the MS are buffered into the old AP during the MS moves and then forwarded to the new AP totally. In addition, because the DSC determines handover successful or not directly, the new AP can send reassociation response without considering reauthentication, resource reservation questions, and so on, which results in the handover time reduction. When an AP finds a MS away and doesn't receive any DsNotify, it sends a Dslnquiry to the DSC. On receiving the Dslnquiry, the DSC searches its database

260

and notifies the AP the MS newest information in DsResponse. The source address of DsResponse is the MAC address of the MS, which results in updating all bridges/switches passed by. Compared with figure 5(b), this method doesn't need the participation of the new AP and controls the traffic reasonable. 3.3 the synchronization of DSC and its database When a DSC fails, its database loses. If the new DSC cannot keep its database same with the former DSC's, some information, such as the locations and QoS requirements of MSs, is nonexistent. Then before recollecting all information, the whole WLAN works abnormally. In the part 3.1, we had described a method to keep the database continuity when a new AP replaces the existing DSC. Here we only modify little on step 3 in figure 6(a) to keep all APs databases synchronization. In step 3, the DSC sends a DsNotify with IAPP multicast as its destination address instead of the old AP's. Then, all APs can receive the frame, know that a certain MS moves to a new AP, and modify their database. So all APs have a unique map of all MSs in the WLAN. In part 3.1, without receives DsBeacon in consecutive 3 DsBeacon interval, DSCCs know that DSC has failed. This method introduces delay. In this period, when a handover occurs, the new AP cannot get information from DSC and the handover cannot complete. We make some improvements. When the new AP send handover request to DSC and doesn't receive its response in consecutive 3 times, it can assume the DSC has failed. Then it can broadcast DsNotify to inquiry the DSC status. The DSCC1 receives it and know that DSC had failed. It will become DSC and sends DsBeacon out. Because all APs have the same database, it's reasonable to regard the replace process is smooth. Then the new AP can send handover to the new DSC. Usually, the ack interval of handover request is much smaller than DsBeacon interval. So we can bring forward the detection of the failure of DSC by using 3 ack intervals instead of 3 DsBeacon intervals and make a new DSC work.

4 Transfer agent In Figure 3, a MS roams from an IP sub-network to a new IP sub-network. Usually the IP mobility can be solved by Mobile IP. In Mobile IP, there are two

261

components: foreign agent (FA), home agent (HA). The FA broadcast advertisement message and HA creates tunnel. The FA and HA are all work in layer 3. By extend it to layer 2, some information, such as SSID, the MAC address of DSC, etc., can be included in FA and HA. In advertisement message, the layer 3 information and layer 2 are all broadcast. LAN1 | R1 | | HA | |

J M B

I FA3 L—r-

DSC1

|

•+*—H—^ | | R3 h ^ " 1

- l-| 1( | DSC3 | | AP3 |

Internet

^^H 1 1

R2

II 1 L

FA2 I

1| AP2 | |

I * DSC2

|

1 MS1 r-^ ~A MS1 I Figure 7. The extended transfer Agent When MSI receives the advertisement message broadcasted by FA3, it knows that it had roamed to LAN3. It encapsulates the IP address of FA3, MAC address of DCS3, and MAC address of AP3 and sends it to HA. HA gets these information and informs DSCl the MAC address of DSC3 and AP3. DSCl knows that MSI had roamed to DSC3/AP3 and modify its database. Then DSCl will inform some layer 2 information of MSI to HA. If MSI is roamed from a foreign network to another foreign network, the DSCl will inform HA the presence of DSC2. HA will inform FA2 and DCS2. DSC2 will delete information of MSI and notice AP2 to delete information of MSI. If AP2 buffers some frames destined to MSI, it can delete them directly or send them to R2. R2 will route them to Rl, then to HA. After receive the information from DSCl, HA will create a tunnel between HA and FA3 and send information of MSI to FA3. This tunnel can be used to transfer packets destined to MSI. FA3 receives the information and sends the relative layer 2 information to DSC3. DSC3 modify its database and ask AP3 send a handover request. AP3 send a Dslnquiry with the MAC address of MSI as the source address and IAPP multicast as the destination address. Then all APs in LAN3 know the location of MSI.

262

By extending transfer agent function from layer3 to layer 2, the layer 2 information can be transferred easily. Then we can solve the scene of roam across IP network. The disorder problem of frames destined to MSI, which occurs when AP2 send buffered frame to Rl, can be solved by MSI. 5

Conclusion

This paper compares 3 available IAPP protocol first, proves the protocol of [3] has the simplest complexity and the least robustness, the protocol of [5] has more complex complexity and a better robustness, the protocol of [4] has the most complex implementation and the best robustness. The flexibility of DSC is also confirmed, which can overcome the possible failure of master AP. Strengthen the role of DSC in central controlled mode is needed. In the IAPP enhancement protocol, the configuration, management and smooth transition of DSC is guaranteed by modify the DsBeacon and the handover process. By extend transfer agent to layer 2, MS's layer2 information can be easily transferred among IP networks. This ensures the smooth roam of MSs when they across IP network.

References 1.

2. 3. 4. 5.

IEEE, "Reference number ISO/IEC 8802-11:1999 (E) IEEE STD 802.11, 1999 edition. International Standard [for] Information Technology-Telecommunications and information exchange between systems-Local and metropolitan area networks-Specific Requirements- Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications," 1999. IEEE802.1 If Workgroup, http://grouper.ieee.Org/groups/802/l 1. Gary Spiess, IAPP handover scenarios, document: IEEE802.11-00/127, May 2000. Gary Spiess, Inter-Access Point Protocol Proposal, document: IEEE802.11-00/345rl, Oct. 2000. Gary Spiess, Inter-Access Point Protocol Enhancements, document: IEEE802.11-00/090, May 2000.

THE HYBRID OF LISTEN-BEFORE-TALK AND ADAPTIVE FREQUENCY HOPPING FOR COEXISTENCE OF BLUETOOTH AND IEEE 802.11 WLAN YONGSUK KIM, BIN ZHEN AND KYUNGHUNIANG, i-Networking Lab, Samsung Advanced Institute of Technology, Samsung, Suwon City, 442-742, Korea E-mail: { yongsuk, zbin, khjang } @samsung.com In Bluetooth system, there are two kinds of interference. One is a frequency static interference, which comes from static interferers like IEEE 802.11 direct sequence. The other is a frequency dynamic interference, which comes from dynamic interferer like other Bluetooth piconets. In this paper we introduce a novel solution of hybrid method of ListenBefore-Talk (LBT) and Adaptive Frequency Hopping (AFH) to address the coexistence of Bluetooth and Direct Sequence of wireless local area network (WLAN). Before any Bluetooth packet transmission, in the turn around time of the current slot, the sender senses the channel whether there is any transmission going on or not. If the channel is busy, packet transmission is withdrawn until another chance. This is the LBT in Bluetooth. Because of asymmetry sense ability of WLAN and Bluetooth, AFH is introduced to combat the left packet collisions. In monitor period of AFH, LBT is performed to label the channels with static interference. Then, all the labeled noisy channels are not used in the followed Bluetooth frequency hopping. In this way, both the frequency dynamic and frequency static interference are effectively mitigated. We evaluate the solution through packet collision analysis and a detail realistic simulation with IP traffic. It turns out that the hybrid method can combat both the frequency dynamic and frequency static interference.

1

Introduction

The 2.4-GHz Industrial, Scientific and Medical (ISM) band is poised for strong growth. Two emerging wireless technologies: Wireless Personal Area Network (WPAN) and Wireless Local Area Network (WLAN) fuel the growth. WLANs, standardized by IEEE 802.11 committee, are designed to cover office or building as an extension of existing wired networks. WLAN devices can operate at bit rate as high as UMb/s and can use DSSS (Direct Sequence Spread Spectrum). The signal symbol is spread with 11-chip Barker sequence and the bandwidth is roughly 22MHz. IEEE 802.11 WLANs access the open air by Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA). The IEEE 802.15 committee is deriving WPAN standard based on Bluetooth foundation specification.'21 Bluetooth is a cable replacement technique, which provides inter-connection of device in range of 10m. It can support bit-rate up to 1 Mb/s for data and voice transmission. A slow FHSS scheme is used at physical level and the hopping frequency covers 79 channels, each channel being lMhz wide. The basic architecture unit of Bluetooth system is piconet, which is composed by a master and up to 7 slaves at most. Each piconet

263

264

master chooses a different hopping sequence. A Time Division Duplex (TDD) technique is used by master and slaves in a piconet. Each packet transmission occurs at the beginning of slot. The attractiveness of working in ISM band is that no license is needed. The downside is that the band must be shared and potential interference tolerated."1 The IEEE 802.11 WLAN and WPAN are complementary techniques. Their devices will likely come together or may come close at a desktop. The interference from 802.11 DS system is termed frequency static interference because they occupy a fixed frequency band. Furthermore, the different piconets may also interfere one another. This is termed frequency dynamic interference. With increased transceiver sensitivity, the number of piconets that may interfere each other increases explosively. Harrtsen simulated the Bluetooth performance in high density 802.11 DS environments and found Bluetooth data link experience more degradation than speech link.[21 S. Zurbes simulated the Bluetooth performance of a large number of co-located piconets with WWW traffic model and found marginally link throughput degrade by interference.131 A. Kameman, N. Ngolmie et al and V. Mitter et al studied the interference of Bluetooth to 802.11 WLAN and vice visa through signal power analysis, detail simulation and experimental measurement.'4"71 Some mechanisms, such as power control, packet scheduling, Time Division Multiple Access (TDMA) of Bluetooth slots and 802.11 slots, Adaptive Frequency Hopping (AFH), and adaptive fragmentation of 802.11 packet were proposed to improve the coexistence of WLANs and WPANs.18"1 ] However, the first TDMA method does not support SCO packet of Bluetooth. The last two schemes need some rules to guide that is the optimal and cannot mitigate the frequency dynamic interference. This paper proposes a novel hybrid method of Listen-Before-Talk and adaptive frequency hopping to avoid the interference from both 802.11 and WPANs. Section 2 introduces the Listen-Before-Talk mechanism and the hybrid method. Section 3 presents an interference analysis and packet collision simulation. Section 4 describes a realistic simulation model and presents the simulation results. Finally, discussion and conclusions are given in Section 5. 2

Listen-Before-Talk in Bluetooth

In Bluetooth piconet, the packet transmission depends on the master's clock and frequency hopping sequences without any consideration of the state of other piconets and channels. The IEEE 802.11 DS system cannot sense the Bluetooth activities, although it senses the channel before transmission. Packet collision occurs when the packets from different networks overlap in time and frequency. The slotted structure of Bluetooth makes it possible to introduce Listen-BeforeTalk (LBT) into Bluetooth. As shown in the upper panel of Figure 1, in any state of Bluetooth Unk controller, there is at least 230us interval before the beginning of

265

packet transmission in next slot. The turn around time is used for Bluetooth device to change from Rx to Tx mode and make the frequency synthesizer tune to the next channel frequency. The lower panel of Figure 1 shows the Bluetooth with LBT, where the dash block with shadow denotes the sense window. Before packet transmission, a check of the next channel, simple energy sense and comparison with a preset threshold, is performed in turn around time of current slot. If the next channel is busy or become busy during the sense window, the sender simply cancels packet transmission, skips the channel, and waits for the next chance. Otherwise, the next slot is free for packet transmission. With a small detection time for channel overlap, the former packet dominates the channel until the end of packet transmission. When packet is withdrawn by sender, the packet sink receive an invalid packet and discard it. f(k)

f(k+l)

Current bluetooth

f(k+2)

turn around time •A-

turn around time

turn around time

listen f(k*2) Listen-beforctalk

listen f(k+3)

listen f(k+l)

Figure 1. Listen-Before-Talk mechanism in Bluetooth

However, there are still some packet collisions between Bluetooth and 802.11 even with LBT. As shown in Figure 2, there are two cases of packet collision, the Bluetooth packet is ahead of the IEEE 802.11 packet and vice versa. Only the latter case of packet collision can be avoided by LBT. This can be contributed to asymmetry sense ability of IEEE 802.11 DS and WPAN. The Bluetooth can sense the activities of IEEE 802.11 DS. However, in general, IEEE 802.11 DS device cannot sense the activities of WPAN.m Compared with IEEE 802.11 DS signal, WPAN signal is narrow band and low power. Bluetooth packet

Q O o

packet collision

WLAN packet

WLAN packet I

u

I bluetooth packet

packet withdraw

Figure 2. Packet collision between IEEE 802.11 DS and Bluetooth

From the above analysis, we find the remained packet collision only comes from the IEEE 802.11 DS packets. As we know, IEEE 802.11 DS signal is static interference.

266

Adaptive Frequency Hopping (AFH) is an effective method for this kind of interference.'91 AFH first monitors communication quality the channel and finds where is the interference. After labeled the channels as 'good' or 'bad', AFH selects a new mapping sequence to avoid the known interference. AFH regularly revert to the original mapping mode to re-evaluates the channels.'91 Therefore, in this paper, we propose a new hybrid method of Listen-BeforeTalk and Adaptive frequency hopping. The LBT senses transient availability of channel before packet transmission to combat the frequency dynamic interference. The AFH monitors long term availability of channel to combat the frequency static interference. If the packet loss or access code error of a channel is above a threshold in monitor period of AFH, it is label as 'bad' channels. After the monitor period, the Bluetooth master and slaves exchange 'bad' channel information and decide the final set of 'bad' channels. In the next stage, all the labeled 'bad' channels are not used in Bluetooth frequency hopping. The 'bad channels' are redeemed regularly after timeout of AFH. 3

Bluetooth interference analysis

In this part, we analyze the packet collision of Bluetooth. We make several assumptions in analysis: • All Bluetooth piconets and IEEE 802.11 WLAN system are independent. • Bluetooth device can sense any packet that appears in the sense window. • Any packet collision incurs packet loss. Figure 3 illustrates the timing of two Bluetooth links. The Bluetooth packets occur on a periodic basis. The Bluetooth packets are transmitted with the period of T, and the transmission data packet sizes are SBi and SBj respectively. Generally, different piconets are asynchronous. Here, we assume the intervals before the start of Bluetooth packet, XBh is uniform distribution as follow (1) XBi~U(0,T) SB,

xBi Bluetooth piconet i —

Ot(ij) XBJ

Bluetooth piconet j _

Figure 3. Timing of two Bluetooth packets from different piconets

n.

267

The probability of Bluetooth packet collision, pdij), is the joint probability of packet overlap in both time and frequency, which is given by = ?*±l*L±

pAitj)

(2)

where C is the number of available frequency channels and C is 79 in most of countries.'11 We denote by OJiJ) the overlap time size. A Bluetooth packet does not collide with more than one packet from the same piconet since tow consecutive packets use different transmission channel. The value of Oc(i,j) depends on packet start position and packet size. We distinguish five cases to compute Oc(i,j) min(S m,SBj) if o = xB,-xBJ min(SB!, max(0, XBj + SBj - X . ) ) Oc(.'J) = min(5 fly ,max(0,X s ,+5 a . - * « » nvu(Q,XB,+SB:-XBJ) msa.(0,XBj+SBj-Xai)

if SBI-SBj<0<XBi-XBj if XB,-XBj<0<SBi-SB] if XBl-XBj<0<SBJ-SB, if $Bj ~ $Bi < 0 < A fl, — A By.

(3)

The packet error probability depends on bit error rate (BER) of physical channel and Forward Error Code (FEC) of the overlap part, as well as p^ij)In all the N piconets, the packet collision probability of a packet from ith piconets is given by Pc(0 = l - flQ-pJiJ))

(4)

We denote by W^j the size of LBT sense window. For the ith piconet, the packet withdraw probability is the probability of sense window overlap with the packet from the any other piconets. Without "hidden terminal", the packet withdraw probability of Bluetooth with LBT, pji), can be written as follow _ * w,„ + SB, i (5) In the same way, the packet collision and packet withdraw between WPAN packets and WLAN packets can be analyzed. In Bluetooth data link, the data packets are followed by an acknowledgement from the recipient in the opposite direction.'21 If we assume all the piconets are active and any packet collision incurs packet error, we can obtain the aggregate throughput of N piconets. For current Bluetooth, the aggregate throughput of is (6) s = (m-*±±l£i»+ F

TC

where S„ and Sd is the packet size of acknowledge packet and data packet, and p is the payload information duty cycle. For Bluetooth with LBT, if we assume all the piconet can hear each other, there will be no packet collision. Then, the aggregate throughput becomes

s^p^a-w^)*-.

(7)

For Bluetooth with AFH, the number of available channel decreases after some channels are labeled as 'bad'. That means the C in Eq.(2) becomes smaller than 79.

268

Figure 4 shows the packet collision and packet withdraw of full size DM1 packet as a function of active piconet number. The simulation agrees very well with the theoretical values. The packet collision and packet withdraw increase with the increment of piconet number. However, the packet withdraw increases much slowly than packet collision because of the smaller size of LBT window. Both packet collision and packet withdraw incurs unsuccessful packet transmission from source to sink. As an example, Figure 5 shows the full size DM5 packet aggregate throughput before and after LBT. For current Bluetooth, the maximal aggregate throughput does not improve when N increases beyond 50, which is about 9.00Mb/s. However, for Bluetooth with LBT, the maximal throughput almost doubles. Curve A: bluetooth

'

.

V* theory o

rtr \ \ J

60

80

100

120

140

piconet number

Figure 4. Packet collision probability and packet withdraw probability of full size DM1 packet as a function of piconet number. The data packet size is 366us, T is 1250ns, and LBT window is 50us.

v simulation

»«-o ^s*>*o

•

.

Curve B: bluetooth with LBT

80

100

120

140

160

180 200

piconet number

Figure 5. Aggregate throughput as a function of active piconet number (DM5 packet). The LBT window size is 50u s, p is 477.8kbps, the data packet is 2862us and ACK packet is 126us.

Figure 6 shows the simulation of aggregate throughput of DM5 packet when there are both IEEE 802.11 DS system and other Bluetooth piconets. In simulation, we assume that the channel monitor process has been finished and all the 'bad' channels have been correctly labeled.191 Due to the interference from IEEE 802.11 DS, the maximal aggregate throughput decreases to about 7.24Mb/s. Compared with the current Bluetooth, the Bluetooth with AFH provides larger throughput only when piconet number is less than 40. The reason is that AFH decreases frequency static interference at expensive of increasing frequency dynamic interference. On the other hand, the Bluetooth with only LBT can only avoid the frequency dynamic interference and partial frequency static interference. It provides more throughput only when piconet number is bigger than 20. For Bluetooth with both LBT and AFH, both types of interference can be avoided. Therefore, it delivers the maximal aggregate throughput.

269

Figure 6. Aggregate throughput with both IEEE 802.11 DS and other piconets interference (DM5 packet). The LBT window size is 50us. For IEEE 802.11, the data and ACK packet size is 1210us and 106us with a lOus interval between them, and the next packet arrives 350us after the end of previous ACK.

4

4.1

IP traffic simulation

simulation model

A more realistic simulation is performed in this section. The considered experiment topology is shown is Figure 7. It is similar to the experiment topology considered by N. Golmie.'81 All the other Bluetooth piconets are placed randomly with uniform distribution in the horizon platform, which can be considered as a rectangular room of size 10m*20m. The slaves are placed around the respective master. The slave device of Bluetooth piconet we concerned locates at coordinates origin, while the master device can move along the horizon axis. Each piconet contains a master and a slave. The IEEE 802.11 DS system locates at the vertical axis, which contains an access point and a mobile device. The access point is fixed at (0,15), while the mobile device is free to move along the vertical axis. The configuration and system parameters are shown in Table 1. We consider a LAN access application. IP packets are generated the same as N. Golmie's simulation.'81 The packet interval is exponentially distributed. We set the offered load to 30% of the channel capacity, which corresponds to mean packet inter arrival times of 29.16ms and 2.52ms for Bluetooth and the IEEE 802.11 systems, respectively. In Bluetooth, the packets are encapsulated with DM5 baseband packets after the corresponding PPP, RFCOMM, and L2CAP packet overheads totaling 17 bytes are added.[S1 For Bluetooth piconet, the IP packet is from master to slave. For IEEE 802.11, the mobile is the generator of data, while the access point is the sink. The IP traffic is organized in session mode, each session is 20s.t8]

270 WLAN access point

(0,15)()

()

WLAN mobile

bluetooth slave / / / /

Other bluetooth piconel device Other bluetooth piconet device

(CM))

(d^O) bluetooth master

/ "

/ /

O

Figure 7. Simulation experiment topology Table 1 Simulation parameters

Bluetooth parameters lmW Transmitted power Slave coordinates (0,0) m Master coordinates (d2,0) m DM5 packet interval 29.16 ms Other piconets distance U( 1,10) m Traffic load 30%

IEEE 802.11 parameters 25 mW Transmitted power (0,15) m AP coordinates (0,dl) m Mobile coordinates 2.52 ms Packet interval 224 bits Packet headers 30% Traffic load Slot time 20 us SIFS 10 us DTPS 50 us CW_min, CW_max 31, 1023

The channel model consists of geometry-based propagation model of the signal, as well as an additive white Gaussian noise (AWGN) model. The AWGN makes the BER to be 10'3 when the distance between Bluetooth devices is 10m. The pass loss for indoor environment is given by fo P SS

°

f32.45 + 201og(/rf) [58.3 + 331og(rf/8)

d
(g)

where/is the frequency in GHz, and d is the distance in meters.14'61 The AWGN, cochannel and weighted adjacent channel 1MHz interference are summed up as noise at receiver. For each received burst, an interference power profile and BER profile are obtained. From BER profile, the access code, header and payload failure probability of each burst is calculated. After bit error detection, error correction is performed according to the FEC algorithm of each part. We assume the LBT energy sense threshold is -70dBm. The LBT window size is set to 50us. Among the methods to monitor the channels, we select a simple one, only the packet loss of each channel is recorded. The AFH first monitors the packet

271

loss of each channel for 4s. In monitor period, any channel whose packet loss is above the threshold is labeled as 'bad' channel. Of course, not all the 'bad' channels can be found in monitor period. After channel monitor, the piconet use a new mapping sequence to avoid 'bad' channels. After the end of each traffic session, the Bluetooth revert to the regular frequency hopping to re-evaluate the channels. 4.2

Simulation results

All simulations were run for 10 sessions. The performance measurements are logged at the slave device of in the concerned piconet. The metrics include packet loss and link throughput. The coordinate of IEEE 802.11 WLAN mobile is fixed at (0, 4) in simulation. Figure 8 shows the packet loss for current Bluetooth, Bluetooth with LBT, Bluetooth with AFH and Bluetooth with the hybrid of LBT and AFH. The distance between slave and master is 2m. IEEE 802.11 DS is an outstanding interferer to Bluetooth. When there are 5 active piconets, among the 0.0854 packet loss, IEEE 802.11 DS contributes 0.071 packet loss. For Bluetooth with LBT, the packet loss is mainly due to the left packet collision with IEEE 802.11 DS packets. For Bluetooth with AFH, the packet loss increases with the increment of active piconets. For the Bluetooth with both LBT and AFH, the packet loss is the least. Figure 9 shows the link throughput when the Bluetooth master-slave distance is 6m, where 'load' denotes the offered traffics. The link throughput decreases with the increment number of active piconets because of more interference. When IEEE 802.11 DS interferer and 5 active Bluetooth piconets co-locate together, only 76.38% offered traffic can be transmitted without retransmission. For the Bluetooth with LBT and AFH, it increases link throughput from 275.06kb/s to 300.84kb/s. 0.14 0.12 ^340

0.1

a

£ 0.08

r

s

320

Link thro

I

EJ LBT

**..

1

! . 0.06

a 300

- V LBT+AH0.04

260

Q„

*""*•

V

LBT+AFH

~"~-o— """©*-. ~ ~ - - - Q bluetooth.

240

0.02 10 15 piconet number

20

25

Figure 8. Packet loss measured at Bluetooth slave device. The master-slave distance is 2m.

piconet number

Figure 9. Bluetooth link throughput of the concerned piconet when master-slave distance is 6m.

272 5

Conclusions

The issue of coexistence is one of the more vexing problems that must be addressed. Bluetooth, IEEE 802.11 WLAN and Microwave oven are likely to be the dominant interferers to Bluetooth in coming year. In conclusion, we propose a novel hybrid method of Listen-Before-Talk and Adaptive Frequency Hopping in Bluetooth to solve the coexistence of IEEE 802.11 DS and Bluetooth. In Bluetooth, before packet transmission in the beginning of next time slot, a check of the next channel is performed in turn around time of current time slot. If the channel is busy during sense window, the sender withdraws packet transmission and wait for next chance. The LBT senses transient availability of channel to combat the frequency dynamic interference. The AFH monitors long term availability of channel to combat the frequency static interference. We observe significant performance gains through packet collision analysis and detail realistic simulation. It turns out that the hybrid method can combat both the frequency dynamic and frequency static interference. References 1. R. C. Braley, I.C. Gifford and R.F. Heile, "Wireless personal area network: an overview of the IEEE P802.16 working group," Mobile computing and Comm. review, vol.4, no. 1, pp.20~27, 1999 2. J. C. Harrtsen, "Bluetooth voice and data performance in 802.11 DS WLAN environment," Bluetooth SIG publication, 1999 3. S. Zurbes, W. Stahl, K. Matheus and J. Harrtsen "Radio network performance of Bluetooth," Proc. Of IEEE ICC2000, vol. 3, pp. 1536-1567 4. A. Kameman, "Coexistence between Bluetooth and IEEE 802.11 CCK: Solution to avoid mutual interference," IEEE 802.11-00/162r0, 2000. 5. N. Golmie, R.E. Van Dyck, and A. Soltanian, "Bluetooth and 802.11b interference: simulation model and system result," IEEE 802.15/195r0, 2001. 6. N. Golmie "Impact of interference on the Bluetooth access control performance: preliminary results" IEEE 802.15-00/322r0, 2000. 7. V. Mitter, I. Howitt and J. Gutierrez, "Empirical Study for IEEE802.11b WLAN & Bluetooth Coexistence in UL Band." IEEE 802.15-01/I48r0, 2001. 8. S. Shellhammer, "Collaborative coexistence mechanism submission: TDMA of 802.11 and Bluetooth," IEEE 802.15-01/025rO, 2000. 9. H. Gan and B. Treister, "Adaptive frequency hopping implementation proposals for IEEE 802.15.1/2 WPAN," IEEE 802.15-00/367r0, 2000. 10. M. B. Shoemake, "Proposal for non-collaborative 802.11 MAC mechanisms for enhancing coexistence: Adaptive fragmentation," IEEE 802.15-01/083, 2001.

A HYBRID ARCHITECTURE OF UMTS AND BLUETOOTH FOR INDOOR WIRELESS/MOBILE COMMUNICATIONS

T. KWON, R. KAPOOR, Y. LEE, M. GERLA UCLA Computer Science, 3803B Boelter Hall, Los Angeles, CA 90095 USA E-mail: {tedkwon, rohitk.yenglee, gerla} @cs. ucla. edu A. ZANELLA Universita degli Studi de Padova, Via Gradenigo 6/A, 35131 Padova, Italia E-mail: [email protected] Over the past several years, mobile/wireless communications have evolved and proliferated a variety of solutions, and in some technologies, standardization is still under way. Among those technologies, we focus on UMTS and Bluetooth because of their prominent and complementary features. While UMTS provides "universal connectivity," Bluetooth will be the most promising candidate technology to provide "ubiquitous connectivity." In this paper, we advocate that a hybrid architecture of UMTS and Bluetooth provide a "one-architecturefits-all" paradigm for any wireless communications. We detail the proposed architecture in indoor environments, where UMTS-TDD is chosen for connecting Bluetooth piconets and scatternets. The key node in this architecture is a "hybrid device" equipped with both UMTS and Bluetooth interface. The performance of the architecture is evaluated by simulation experiments considering the number of Bluetooth devices per hybrid device under diverse configurations.

1

Introduction

Over the past several years, there have been revolutionary changes in the communication world. One trend is the proliferation of mobile/wireless communication technologies. The 2nd generation mobile cellular networks can be regarded as the flagship of that trend, which focused on voice telephony services. Another significant current is ever increasing data traffic due to the Internet. The need for untethered Internet brings forth to diverse wireless LAN solutions such as IEEE 802.11 [4] and Bluetooth [6]. The 2nd generation cellular communication technologies inherently aim at voice services, so that bursty Internet traffic is not readily supported. Furthermore, there is little compatibility among different countries and even among different cellular service providers in the same country. Therefore, the 3 rd generation cellular communication technology is currently standardized with slogan of "universal connectivity" and "higher bit rate" for the bursty Internet traffic. Universal mobile telecommunication system (UMTS) is the 3rd generation cellular system that is standardized by 3GPP [1]. In this paper, we adopt UMTS because of "universal

273

274

connectivity," which is essential for the paradigm of "one-architecture-fits-all" in any wireless communication environments. Another rationale of choosing UMTS is that almost every user is likely to carry UMTS mobile terminal (or "user equipment" in 3 GPP documents) in the near future. Even though UMTS has the advantage of global coverage and roaming, its limitation is that wireless communication is possible only between UMTS base station (BS) and UMTS user equipment (UE). The other technology that complements UMTS is Bluetooth. Bluetooth is a short-range, low power, and low cost technology, which is expected to be widely diffused to almost every electronic device. That is, Bluetooth will provide "ubiquitous connectivity" among any devices only if they have Bluetooth interface. One of the most important feature of Bluetooth is "ad hoc network connectivity" by which any devices equipped with Bluetooth can formulate a piconet (a basic single hop network of up to eight Bluetooth devices) and a scattemet (network of piconets). By this "ad hoc network connectivity" and "ubiquitous connectivity," Bluetooth will fill up the shortage of UMTS. From the standpoint of Bluetooth, its limited radio coverage range and the small number of devices in a piconet represent severe constraints in realizing pureBluetooth wireless networking solutions. For example, as the number of Bluetooth devices increasing, the size of scattemet has to be increased, which, in turn, the number of hops between communicating Bluetooth devices may become larger. Unfortunately, as the number of hops in a scattemet increases, the performance of the scattemet degrades rapidly. Thus, there should be a limit on the size of a scattemet to be practical wireless communication solutions. One way to mitigate this scalability problem is using other technology between scatternets. For example, the scattemet can then be connected to other scatternets via UMTS. Some may argue that in indoor environments, IEEE 802.11 be better technologies to connect Bluetooth scatternets, especially considering higher bandwidth capacity. However, the current IEEE 802.11b shares the same ISM frequency band with Bluetooth, so that there is an interference problem [5]. Also, it is highly likely that a user will always carry an UMTS UE rather than a device with 802.11 interface. Furthermore, some chips in cellular handset (e.g., QUALCOMM's some MSM series) have already incorporated Bluetooth interface, which makes the proposed architecture easily deployable. Especially in indoor environments, UMTS radio coverage may not be in a favorable condition and in some cases, an UMTS UE may not even catch the radio signal. In that case, an UMTS UE that has a Bluetooth interface can be accommodated by Bluetooth access point (AP), which provides a connection to a wired Internet. Taking the above into account, we propose a new hybrid architecture of UMTS and Bluetooth, where interworking of two networks is performed by a "hybrid device" that has both UMTS and Bluetooth interface. We believe that the most

275

popular form of the "hybrid device" will be the UMTS UE that has a Bluetooth interface. The rest of this paper is organized as follows. Section 2 describes technical aspects of UMTS and Bluetooth technologies. The proposed architecture and issues therein will be presented in Section 3. Numerical results in indoor environments are shown in Section 4. Section 5 concludes this paper. 2

2.1

Technical Overview

UMTS

UMTS is the 3 r generation mobile communications system being standardized within IMT-2000 framework. UMTS will build on and extend the current 2nd generation cellular technology by providing increased capacity, a variety of service, and accommodation of data traffic. This enhancement can be achieved by the innovative radio transmission technology and the evolving core network. There are mainly two air interface technologies for UMTS: UMTS-FDD and UMTS-TDD. UMTS-FDD relies on wideband-CDMA (W-CDMA) technology and will be deployed in outdoor macro-cellular or micro-cellular communication environments. UMTS-TDD adopts a combination of CDMA and TDMA technologies and will be deployed in indoor pico-cellular communication environments. It allows asymmetric radio resource allocation between uplink and downlink and higher bit rates services than FDD and is expected to be deployed in indoor environments and hot spots. Here, we think of UMTS-TDD as the candidate technology that can play an essential role in indoor wireless/mobile communications. 2.2

Bluetooth

Bluetooth is conceived both as a simple personal area network (PAN) cable replacement technology and as a universal low cost wireless connection into both voice and data networks. Bluetooth is specifically designed for low power, low cost operation and will be built in as standard into PC, cellular telephones, printers, PDA's, fax machines, keyboards, joysticks and even electronic appliances. Bluetooth operates in the 2.4GHz ISM frequency band uses a fast frequencyhopping technique to minimize interference. A Bluetooth unit has a range of approximately 10 meters. Two or more Bluetooth units sharing the same channel form a piconet. Each piconet consists of a master unit and up to seven active slave units. Furthermore, two or more piconets can be interconnected to form a scatternet.

276

This requires a unit, called an inter-piconet unit (gateway), to be a part of more than one piconet. 3 3.1

The Hybrid Architecture An Integrated Scenario

We illustrate the proposed architecture as shown in Fig. 1, using a hierarchical approach. At the top level, there is an UMTS terrestrial radio access network (UTRAN), which consists of an UMTS base station (BS) and UMTS UEs. Recall

! >

-t

r

1

I / O

,

•""^

T -*h . n

r

UMTS •

^—^r-

Scattcrnet Piconet

(

!

••"o

< \

h A

I Ivbrid

-r^

Master •

Gateway 0

slave unit

Fig. 1. A Hybrid Architecture of Bluetooth and UMTS

that UMTS UEs are hybrid devices that also have a Bluetooth interface. Therefore, the hybrid devices deal with interworking between UTRAN and Bluetooth networks at lower levels. At the second level, there are scatternets, each of which contains at least one hybrid device. Each scatternet consists of one or more piconets and piconets are inter-connected by a Bluetooth gateway. Note that the second level in Fig. 1 depicts only logical grouping of piconets in a scatternet, not the physical network. At the bottom level, there are piconets, each of which consists of one master and up to seven slave units. One of the slave units is the Bluetooth gateway. The key element of the proposed architecture is the hybrid device. The hybrid device can operate with both interfaces simultaneously. The routing layer of the

277

hybrid device should decide to which interface packets are to be forwarded. In Fig. 1, packets in a scattemet may be aggregated towards the hybrid device if the destination of packets is outside the scattemet. The packets then forwarded to the UMTS BS from the hybrid device. Let us discuss an example topology of a scattemet in Fig. 2. This scattemet consists of 3X3 piconets where the hybrid device is located in central piconet. The gray lines in the figure show a possible way in which packets in each piconet are forwarded to the hybrid device. Such a topology may be applied in many different scenarios. For instance, in an "Intelligent-Supermarket," a central server (it can be connected to any node in the Fig. 1, say, the UMTS BS or a Bluetooth unit.) may contain information about each subscribed client - the usual grocery list, the kind of offers he/she may be interested in, his/her account information (e.g., credit card number). When the client enters the supermarket, his/her identification code is sent to the server through the above ® Hybrid \ j # Master •

Gateway

Fig. 2. An example topology of the architecture showing a 3x3 Bluetooth scattemet

wireless network (by means of service discovery). The server can, then, send back different messages that may direct the client towards the products he/she is interested in, or advise him/her about "special offers." The messages may be displayed on the client's cell phone or palmtop or, perhaps, on a little screen applied to the market-cart. The master node in each piconet may correspond to the Bluetooth unit located at each category of products such as meat, vegetable, and so on. 3.2

Fairness Issue

Although the connectivity can be readily dealt with the above architecture, there is a fairness issue in sharing bandwidth capacity. The fairness issue can be happening at top level (UTRAN) and at the second level (scattemet) in Fig. 1. At top level, UMTS-TDD BS will allocate the bandwidth to the hybrid devices

278

according to the request from the UMTS UEs (hybrid devices in the proposed architecture) or the queue length in the BS. More specifically, UMTS BS first divides the bandwidth capacity between downlink and uplink. Then, packets to the individual hybrid devices will be transmitted using downlink bandwidth and packets to the UMTS BS from the hybrid devices will be transmitted using uplink bandwidth. This bandwidth allocation is performed by radio resource control (RRC) entity in UMTS control plane. As there can be deviation in the number of active Bluetooth units per the hybrid devices, the RRC should accommodate the diverse scenario of non-uniform traffic distribution. Fairness at the second level (scatternet level) is more complicated in that there are different number of hops (piconets) between Bluetooth units and the hybrid device in the same scatternet as shown in Fig. 2. Suppose every Bluetooth unit has packets to be forwarded to the hybrid device. Whenever the packet is forwarded between piconets in the given scatternet, this forwarding is handled by the gateway

Fig. 3. Part of a scatternet showing the number of users (slaves) in the piconets

unit. Since a gateway unit can be present only in one piconet at a time, the total capacity of a gateway can provide to the users (Bluetooth units in that piconet) is bounded by half the piconet capacity. Fig. 3 shows an example of a part of a scatternet along with the number of users in some piconets. As before, the gray lines show the multi-hop paths between the hybrid unit and the other piconets. Note that piconets whose number of users is shown are served by the same gateway Gi of the central piconet. Let the total number of users in the scatternet be 25 (only part of them are shown). Then, the gateway Gi serves a fraction 15/25 of the total number of users and hence, spends a fraction 0.5 of its time in the central piconet. It divides its remaining time between the other two piconets, giving each a forwarding fraction of 0.3 and 0.2 of its total time, respectively (since the piconets serve a total of 9 and 6 slaves respectively). This fair division requires a coordination of gateways and masters that may be

279 implemented as controlling the polling cycles of masters of piconets Pi and P2. Refer to [10] for the detailed discussion. 4

Numerical Results

We used GloMoSim [8], a scalable simulation library, to develop both simulation models: Bluetooth model and UMTS model. We also integrated the two simulation models to a hybrid device of Bluetooth and UMTS. The Bluetooth baseband and L2CAP layer was implemented according to the specifications [6]. The Time Division Duplex (TDD) slot length in Bluetooth is equal to 625 microseconds. The master transmission starts in even numbered slots

1400 1200 .—. Q. .Q J£

1000

800 3

£n D) 3 O .C

1-

HOI)

400 200 _ B _ D = 16 —H— D = 1 6 0 —A— D = 800

0 ~f"*1 2

,

3

.

r—

4

,

5

— ,

6

—

7

,

r-

8

7~""

9

"1

10

Average Interval (s) Fig. 4. Throughput of two hybrid devices vs. interval between FTP sessions

and slave transmission starts in odd-numbered slots. Packets can be 1, 3, 5 slots long and are transmitted in consecutive slots. A different hop frequency is used for each packet. A slave is allowed to transmit in a given slot if the master has addressed it in the preceding slot. For each slave, there is a corresponding queue at the master. The synchronization process is assumed to be complete. When a gateway node leaves, it informs the current master that it will be unavailable for predetermined period. During its absence, the master stops to poll the gateway, and the traffic on the piconet between the master and other slaves/gateways continues as usual. The UMTS-TDD simulator was developed according to the specifications [1]. In the radio layer, there are 12 slots in a frame reserved for data, and the other 3 slots for downlink and uplink signaling. There can be up to 16 dedicated channels in each slot for users' data. Also, we adopt the turbo coding with rate 1/3 for forward

280

error correction (FEC) and quadrature phase shift keying (QPSK) for modulation. As for MAC layer, we assume the selective reject scheme for error control. We adopt a dynamic radio resource allocation algorithm, where the radio resource units are allocated to individual mobile terminals considering their queue length. We run three sets of experiments, where the routing protocol used is AODV [9] and each experiment is run for 400 seconds (in simulation time). Throughout the whole three sets of experiments, there are always six Bluetooth devices which download the files from the UMTS BS. In the first set of experiments, we consider a scenario where there are two hybrid units and each hybrid unit is the master of 3 Bluetooth devices, each of which downloads TCP traffics of 100KB from the UMTS BS repeatedly. The 100 KB file is downloaded repeatedly with fixed intervals as will be shown in X-axis. Each device, moreover, is the master of its own piconet, connecting other 4 slaves that do not generate traffic, but get polled

1200 1100 a. 1000 900 3 800 O. SZ 700 01 3 600 O 500 400 0

a

D = 800*8 4 6 8 Average Interval(s)

10

12

Fig. 5. Throughput of six hybrid devices vs. interval between FTP sessions.

periodically with a DHl packet by the master. The polling cycle at each Bluetooth device for its slave units is D slots. During each polling cycle, the device must poll each its slave once and listens to the hybrid in the remaining time. The hybrid devices adopt a Limited Exhaustive Round Robin polling scheme. In other words, the hybrid unit polls each Bluetooth device until it has packets to exchange, but not for more than X slots (X < D/3). The value of D can be decided on the basis of the maximum delay that the slaves can accept. We consider three different values of X: 5 slots (D=16 slots), 53 slots (D = 160 slots), and 266 slots (D= 800 slots). Fig. 4 shows the results from the first set of experiment where the X-axis is the interval between successive FTP sessions. Note that as the interval between FTP file download increases, the traffic volume transfer rate decreases. We can observe that as D decreases, each Bluetooth device wastes its piconet bandwidth capacity more

281

because it has to poll its slaves more frequently. This phenomenon becomes clear when the interval between FTP sessions is short. In the second set of experiments, there are six hybrid devices and six Bluetooth devices, so that each hybrid device needs to service only one Bluetooth device. The other parameters are the same as the first set. Fig 5. shows throughput of the second set of experiment vs. the interval between FTP sessions. Here, there is little difference in spite of different D values because of there is only one Bluetooth device per hybrid device. The bandwidth wasted for polling does not have a significant effect in this case. Fig. 6 shows the third set of experiments with three hybrid devices and six Bluetooth devices for FTP sessions. There are therefore two Bluetooth devices per hybrid device. In this configuration, there is a loss of throughput only in the case of D = 16 slots, where polling of slave units wastes the piconet bandwidth most. This observation also verifies the above reasoning. 1400 _1200 Ass*,. 8-1000 -

^

3

* V

B

H

—--:=:^===----_ = 5

2 400 H

=aaa

B

—

200 - B — D=16 0 I

2

,

3

4

- H — D-1G0 ...„,

-A— D-800 '

,

5 6 7 Average Interval(s)

,^.,p

8

9

.

10

Fig. 6. Throughput of three hybrid devices vs. interval between FTP

5

Conclusions

Although there have been proliferated a number of wireless/mobile communication technologies, each technology has its own strong points and weak points in terms of radio coverage, bandwidth capacity, mobility-support, and so on. In this paper, we presented a hybrid architecture of UMTS and Bluetooth that can provide a "onearchitecture-fits-all" solution for any wireless communication environments. The key element of the architecture is the "hybrid device" equipped with both UMTS interface and Bluetooth interface. We illustrated the possible configuration of the hybrid architecture in indoor environments, discussed the fairness issue, and presented the possible scenario of "intelligent supermarket," where the users can download the information from central server via the network of the proposed

282

architecture. We evaluated the hybrid architecture by simulation experiments with varying number of Bluetooth devices per hybrid device. As the number of Bluetooth devices per hybrid device decreases, the polling frequency (with which Bluetooth device polls its own slave nodes) has less impact on the throughput of the overall system. References 1. ETSI TR 125 922 V3.5.0 (2001-03) Universal Mobile Telecommunications System (UMTS), http://www.3gpp.org/3G_Specs/3G_Specs.htm. 2. J. C. Haartsen, "Bluetooth: A New Radio Interface Providing Ubiquitous Connectivity", IEEE VTC 2000 - Spring, pp. 107-111. 3. C. Gessner and R. Kohn, J. Schniedenharn, and A. Sitte, "Layer 2 and Layer 3 of UMTS-TDD," IEEE VTC 2000 Spring, pp. 1181 -1185, Tokyo, 2000. 4. Specifications of IEEE 802.11, http://standards.ieee.Org/getieee802/802.l 1.html 5. C. F. Chiasserini, R. R. Rao, "Performance of IEEE 802.11 WLANs in a Bluetooth environment", Wireless Communications and Networking Conference, WCNC 2000. 6. Specifications of the Bluetooth System - Core vol. 1 vl. 1, www.bluetooth.com 7. C. Mihailescu, X. Lagrange, and Ph. Godlewski, "Dynamic Resource Allocation for Packet Transmission in UMTS TDD TD-CDMA Systems", IEEE Vehicular Technology Conference (VTC), Houston, USA, May 1999. 8. X Zeng, R. Bagrodia, and M. Gerla. "GloMoSim: a library for parallel simulation of large-scale wireless networks", In Proceedings of the 12th Workshop on Parallel and Distributed Simulations, May 1998. 9. C. Perkins, E. Royer, "Ad-hoc on-demand distance vector routing", IEEE Workshop on Mobile Computing Systems and Applications (WMCSA '99), Page(s): 90-100. 1999. 10. M. Gerla, Y. Lee, R. Kapoor, T. Kwon, and A. Zanella, "UMTS-TDD: A Solution for Internetworking Bluetooth Piconets in Indoor Environments," submitted to WCNC 2002.

Part 10 Multicarrier Systems

ADAPTIVE LINK ADAPTATION FOR MULTICARRIER SYSTEMS FEI TANG: STEVEN THOEN+MARC ENGELS AND LUC DENEIRE Interuniversity Microelectronics Center (IMEC) Kapeldreef75, B-3001, Leuven, Belgium E-mail: [email protected]

Adaptive Loading (AL) is a successful technique to significantly increase the spectral efficiency (or minimize BER) of multicarrier systems. However the complexity of existing AL algorithms is rather high. Moreover, we still want to increase the spectral efficiency further. Therefore we apply a new technique called Adaptive Link Adaptation (ALA) in a multicarrier system (HIPERLAN/2) to address these two problems. This paper studies the feasibility and performance of such a technique in HIPERLAN/2. Simulation results show a 3.5 dB gain over the existing AL algorithm with the best performance. More importantly, the complexity is greatly reduced roughly by a factor of 100, which facilitates the implementation.

1

Introduction

The research at IMEC already has given birth to a second version of an ASIC core for an OFDM modem implementing the physical layer of the HIPERLAN/2 and IEEE 802.11a systems '. In order to meet the requirements of next generation WLANs, we have to further improve the spectral efficiency. This improvement can be achieved by Adaptive Loading (AL), a promising technique that takes advantage of the variation of SNR across different carriers and tries to allocate power and bits in an optimal way. The major existing AL algorithms, Hughes-Hartogs's 2 , Chow's 3 and Fischer's 4 loading algorithms, only use the information of frequency-domain variation, while wireless channels are time varying. This time variation can be exploited to further improve the spectral efficiency. Moreover, the implementation cost of previous algorithms is very high, especially for the Hughes-Hartogs algorithm, which makes the hardware implementation difficult. The paper studies the feasibility of a new technique called Adaptive Link Adaptation (ALA) 5 (which was previously studied for the single carrier system) in the HIPERLAN/2 system, quantifies the possible additional gain, and evaluates the complexity. Simulations of ALA against the three major AL algorithms in the HIPERLAN/2 channel model show a gain of at least 3.5 dB. And more importantly, the complexity of ALA is reduced dramatically.

* ALSO A PH.D. STUDENT AT E.E. DEPT., ESAT/INSYS, K.U.LEUVEN tFUNDED BY THE FWO

285

286

-

1

1

1

1

1

-6-7-8

-10

1

6 7 8

—

—

—

-20 -21 -22

DC — Pilot Data

Q.

0.

Q.

-26

P

DC

—

—

20 21 22

26

Figure 1. HIPERLAN/2 Carrier Frequency Allocation

2

HIPERLAN/2 Channel Model

The HIPERLAN/2 reference model 6 consists of a tapped delay line. There are 18 non-uniform distributed taps. Except for the first tap, which can have the Ricean fading statistics, each tap has the Rayleigh fading statistics. A classical Doppler spectrum corresponding to the speed of 3m/s (about 50 Hz in 5G Hz band) is assumed for all taps. We use model A, which has the largest delay of 390 ns and all taps have a Rayleigh distribution. In the HIPERLAN/2 system standard 7 , there are totally N = 64 carriers. However not all carriers are used. Only Nj = 48 data carriers and Np = 4 pilot channels are used. Fig. 1 shows the specific distribution of these 52 carriers. The sampling period is 50 ns, so the original DMT symbol duration is Ty = 50 x 64 = 3.2 /is. After adding the cyclic prefix of 800 ns, the total symbol duration is 4.0 jus. Thanks to this cyclic prefix, which is chosen to be larger than the channel's impulse response, it can be proven that the equivalent channel for each carrier is simply a complex value H(iAf) while H is the DFT of current time domain channel response h. And it can be further proven that each needed DFT value is simply a linear combination of the 18 current taps' values. This can be represented as H — A x tap where H is a 64 x 1 column vector, tap is a 18 x 1 column vector, and A is a 64 x 18 matrix. Each element of matrix A has an amplitude of 1. Since each tap has the Rayleigh statistics, tap is a complex Gaussian process vector. As each element is independent of others, the amplitude of H is also a Rayleigh process vector and each element has the same average energy Ec, which is equal to the sum of original taps' energy. The normalized correlation matrix between different carriers is ^? =

A x TAP x A' Ec

, where TAP = [tap, tap,•••, "

v N=64

tap].

(1)

'

In practice, we can keep some rows of matrix A to generate H only for the need

287

Figure 2. Correlation Matrix of the 48 data Carriers

carriers. Fig. 2 depicts the amplitude of the normalized correlation between the 48 data carriers.a 3

ALA for HIPERLAN/2

We define the channel quality q(t,f), Signal to Interference and Noise Ratio (SINR), and the channel capacity Cmc of a wide-band channel between frequency / I and / 2 , timerl and 12 as:

q{tj)

a(t,f) SINR(t,f)=P(t,f)q(t,f), Na + I(t,f)' 2

(-*mc _mc -—

f max Jt\ P(l,f):jfP(t,f)dtdf=E.

f

log2(l +P(t,f)q(t,f))dtdf,

(2)

(3)

Jf\ where a is the fading energy, Na is the additive noise and / is the possible interference which has the same property as the additive noise, E is the available energy, P(t,f) is the power adaptation function. Similar to the one-dimensional case 8 , it is easy to prove that the optimal solution is a two-dimensional "water-pouring" filtering along a

Another method to generate individual channels is using RxG, where G is a vector whose elements are independent complex Gaussian processes. The elements' complex modulus has the Rayieigh statistics and is shaped by the Doppler Spectrum.

288

both the time and frequency domain. The power and bits are variable both in time and frequency. Existing AL algorithms checked with the "variable power, variable bits" principle is shown is Table 1. For the time-domain, we will use the distribution of the channel quality q(f), which contains all lst-order time-varying information. Supposing we have several coding and modulation schemes, we can characterize each combination by the required SINR thresholds th respectively, and select M best working modes from those combinations by evaluating rates R against the corresponding th values. This can be expressed as: th(i)
R(i)
i = 0,---,M-l,

(4)

J?(0) = 0.

(5)

For such a practical system, the simple "water pouring" filtering does not work since 1) the rates are discrete; 2) the number of carrier is finite; 3) each mode can not work at the channel capacity. However, for optimality, it remains necessary to allocate more bits where the channel condition is good. For simplicity, we first assume that there is only one flat-fading carrier. It is easy to formulate the following optimization problem to maximize throughput with an average power budget P/, while maintaining a constant BER. M

Ati+l

/:maxX/ x

i=\Jxi

R(i)p(q)dq,

(6)

subject to g\ g2,-,M+l

• 0 < X\ <X2-Xi<

Xi+i

••• <XM-

(8)

Different modes with rate R(i) correspond to different non-overlapping floating regions defined by the boundaries *,. If q falls between x; and x-l+\ (i — 0, • • •, A/ 1, xM+i = +°°), the system will work in mode i. Within each region, the power may be adapted to maintain the SINR P x q constant at the threshold th{i). The global solution to this problem is 9 : Xi=tct,

withc, = f ; . J

*„,.

J,

i=l,---,N,

t>0.

(9)

Substituting JC, into (7), we can solve the one-variable nonlinear equation to get the value of parameter t. For the multicarrier case with N flat-fading carriers, we only need to replace p(q) and Pb in (6) and (7) with Pmc{q)=^Nn=l^{q\

and

P-^Pb/N,

(10)

289

where pn{q) is the channel quality distribution of subcarrier n. Here, for a certain moment in time, we unfold the frequency axis, and pack corresponding N q values into this moment. With this adaptation, our previous results for single carrier can be directly applied. So the optimal switching boundaries *, have a simple linear relationship regardless of channels, modulation and coding. They are all proportional to t with a constant c,- which is equal to the threshold difference over rate difference of neighboring modes. For the HIPERLAN/2 channel model A, since pn{^/q),i = ", • • •,N are identically Rayleigh distributed (when no interference exists), we have a closed-form for Pmc{q), which is the same as pn(q)- However, when the cyclic prefix is shorter than the maximum delay, each carrier experiences InterSymbol Interference (ISI) and InterCarrier Interference (ICI) and some kind of equalizer is required. b In this case, we are not be able to find a closed form for pmc. In realistic circumstances, we may not even know pmc. Moreover, the channel can be non-stationary as interference for some carriers fades in and fades out and the statistics may change over time. So we prefer to arrive at the correct value of f on the fly and thus adapt the adaptation scheme per se. We term this technique as Adaptive Link Adaptation. The ALA system is initialized with a t value for a typical indoor channel and will do the normal resource allocation. But it will update the adaptation scheme iteratively c if the channel assumption is not perfect or whenever the channel statistics changes. The actually used power for a carrier at a certain moment is P = th(i)/q. From (7), we can derive

g-i*'-',)-"('W dt

^

a.,

t

Thus we obtain the relationship of small variation of t and P at the optimal point, i.e. (12) K At = sAP, with s = £ —i 1 —p(ci t) In practice, we will use a variable step size s instead. It has a large value when AP is large so as to speed up the convergence process. But it has a small value when near the solution so as not to alter significantly the almost correct t value. s = -jS\AP\m; f„ = r „ _ 1 - j 3 | A P r + 1 , in = 2, 0 = 0.3.

(13) (14)

*In this case, we need an extended channel which includes the equalizer. However, we assume the extended channel is still linear as the above results require a linear channel. T o r the proof of convergence at the global solution, see 5 .

290 So the actually used power is first low-passed to filter out the DC component. At iteration n, AP is obtained by subtracting the DC component from P'b, and AP is multiplied by the step-size s to adapt t, and hence the adaptation scheme. In the single carrier case 5 , using simulation as a feedback, we choose P values of 10 periods of Doppler frequency to average out, in order to reliably estimate the DC component. From Fig. 2, we know that the absolute value of normalized correlation between two carriers will be very small, approximately below 0.2 and 0.04 in energy when the distance is larger than 6 carriers. As at a certain moment, we have N values of P from N carriers which are not highly correlated, we need less periods to extract the DC component. Likewise, using simulation results as a feedback, we choose \2/N periods of power values of all carriers, which will speed up the convergence process. 4 4.1

Simulation Results and Comparison Performance and Comparisons

We assume the simplest scenario such that the theoretical analysis is possible. MQAM with constellation size of 2,4,16,64 is used. HIPERLAN/2 channel Model A is used with additive white noise. P'b is normalized to 1. We try to do a fair comparison of heterogeneous schemes which have different constraints and optimization goals. The BER constraint is set to 10 - 3 when applicable. For Chow's algorithm, we will find simulation points correspoding to BER of 10~3. The number of loaded bits nomalized by the number of carriers against average SNR across all carriers d is used as the performance measure. And the prefix overhead of 4/3.2 = 0.97 dB is also taken into account. Two numbers of carriers are used. In one case, we use the original 48 data carrier, and in another case, we select the 20 carriers from the center of the 48 carriers. The simulation results for all schemes are shown in Fig. 3 and Table 1. For ALA, simulation results coincide with the theoretical results. There is no performance change when the number of carriers is reduced to 20. "Bits/carrier" increase smoothly with the increasing SNR. Hughes-Hartogs's algorithm ranks 2nd in performance with a distance of 3.5 dB to ALA's when "bits/carrier" is 2. But it deteriorates slightly when the number of carriers is reduced to 20. "Bits/carrier" also increases smoothly with the increasing SNR. The underlying logic of HughesHartogs's algorithm is to allocate bit by bit with the lowest required additional power. Although the rate in time-domain is varying, it's not done in an optimal way by considering the time-domain varying information. Also, it uses a constant average power in the time-domain and incurs a small additional loss. The Chow's algorithm rf

More strictly speaking, should be average channel quality q across all carriers scaled by the average used power per carrier, since in this case, E(P*q) is slightly different from E(P)E(q).

291 6r-.

.

.

.

(a) With 48 Carriers

.

.-i

6|-i-

(b) With 20 Carriers

Figure 3. Comparison of Performance in the HIPERLAN/2 Channel

does not perform well in the HIPERLAN/2 scenario. Unlike in DSL, the number of carrier is small and more importantly the allowed possible bits on each carrier is restricted to set {1, 2, 4, 6}. It has to quantize the calculated continuous rate with the restricted set. When SNR is low, the number of "good" carriers is small. So when "bits/carrier" is increased, it can not handle the additional bit burden well with limited choices. However, as the SNR continues to increase, almost all carriers will become "good" carriers, and the sensitivity to the limited choice becomes less (just use the highest rate when possible). As a result, it is expected that the performance of all algorithm will converge when SNR is 4-°°. For Chow's algorithm, once the bit allocation is decided, it is the only one which will use as much power as nessesary to maintain a certain BER on each selected carrier. Due to the limited choices, the otherwisely unused carriers may be used. This results in a large power overhead when some of the carrier have a very low channel quality. Fischer's algorithm has a performance difference of 9 dB at Ibitjcarrier. "Bits/carrier" also increases smoothly with the increasing SNR. However it deteriorates most when the carrier number is reduced to 20. ALA has the lowest complexity. For each carrier, we can use several comparison operations or simply a look-up-table operation, which can be implemented as combinatorial logic. If we choose to use comparison operations, it can be further simplified by comparing with the two boundaries of the last working mode first. Of the two boundaries, the one in the direction of mode change trend is checked first. The com-

292 Table 1. Comparison of ALA and Existing AL Schemes Variable Rate in Freq. Variable Rate in Time Variable power in Freq. Variable power in Time Constraints Goal Coding Support SNR(dB)@2bits/carrier HIPERLAN/2 48carriers SNR(dB)@2bits/carrier HIPERLAN/2 20carriers Log Function Evaluation Number of Operation (N carriers, R bits) Complexity/Operation

Matlab Profiling Time (s) (several bit loads, 1000 DMT symbols per time) HIPERLAN/2,48carriers

Chow Y N Y Y BER Rate Min Power (Implicit) unclear

Fischer Y N N N Power Rate Min BER

22.7

21.5

16.1

23.3

22.7

N

N

Y

Y

0{N)

0{N x R)

Very Low

High

0(Nx runtimecount) Medium

12

2138

O(moxcount XN+2N) High with Log Function Evaluation 1623

ALA Y Y Y Y Power BER Max Rate

Hughes-Hartogs Y Y Y N Power BER Max Rate

Y 12.5

Y (Modified Alg.) 16

12.5

unclear

1545

plexity of update is also low, since we can choose the update process to be triggered only if the used power deviation from the budget exceeds a pre-defined threshold. In the event of an update, for each iteration, it only need simple operation to collect power values to average out and update t value accordingly. As an analytical evaluation of the complexity of all schemes is very difficult, we choose to use simulation to profile them. The most accurate comparison should be done with the cycle-true Register Transfer Level (RTL) VHDL simulation. However, it is still related to the amount of hardware resource used and a fair comparison is difficult. For simplicity, we can get a very rough estimation by profiling the fine-tuned MATLAB codes and summing up the results obtained in different scenarios. As shown in Table 1, ALA performs the best with dramatically reduced complexity. However, Chow's and Fischer's algorithms did not do a lot better as expected than the "complexity-notorious" Hughes-Hartogs's scheme. The reason is that these two algorithms involves lots of decision making and log function evaluation which MATLAB performs slowly. On the other hand, Hughes-Hartogs's algorithm can be represented in an elegant matrix format, which can be optimally executed in MATLAB.

Average SNR in HIPERLAN/2 (dB)

(a) Parameter t value

Iteration Number

(b) Convergence Characteristic

Figure 4. Update of ALA in HIPERLAN/2

4.2

Convergence of ALA in HIPERLAN/2

If we change the first tap of HIPERLAN/2 channel model into Ricean fading and follow the same derivation in Section 2, we can prove that each subchannel is a Ricean fading channel though with values of average power and Ricean A: factor different from those of the 1st tap. As the Ricean distribution is the same for all carriers, pmc{^) will be Ricean distributed if no interference exists. By altering the fading property of the 1st tap, we can have pmc{<j) of various distributions. Fig. 4(a) shows the numerically calculated theoretical values of parameter t under various conditions. (The prefix overhead is still included.) Fig. 4(b) shows the convergence process when pmc{y/q) changes from Rayleigh fading with average SNR of 8 dB to Ricean fading with Ricean factor k = 6 dB and average SNR of 20 dB using the method outlined in Section 3. From Fig. 4(a), we know that the value of t is supposed to be adapted from 0.66 to 1.42. The adaptation works well and actually for the HIPERLAN/2 multicarrier system, the convergence process is improved by a factor of (10 x N)/12 when compared to similar single carrier systems. 5

Conclusions and Future Work

This paper studies the application of ALA in a multicarrier system HIPERLAN/2. The additional gain, which can be achieved by considering also the channel variation in time-domain, is at least 3.5 dB for a typical HIPERLAN/2 scenario. Furthermore,

294 a comparison with major AL algorithms is given. The most important benefit of ALA is that the complexity is greatly reduced. However, once the channel statistics change, the convergence time is still not very satisfactory. Currently, we are experimenting and refining an improved update algorithm for ALA, which may speed up this process. References 1. W. Eberle, M. Badaroglu, V. Derudder, S. Thoen, P. Vandenameele, L. Van der Perre, M. Vergara, B. Gyselinckx, M. Engels, and I. Bolsen. A digital 80 Mb/s OFDM transceiver IC for wireless LAN in the 5 GHz band. In Proc. IEEE ISSCC'2000, Feb 2000. 2. J.A.C. Bingham. Multicarrier modulation for data transmission: An idea whose time has come. IEEE Commun. Mag., 28(5):5-14, May 1990. 3. P.S. Chow, J.M. Cioffi, and J.C. Bingham. A practical discrete multitone transceiver loading algorithm for data transmission over spectrally shanped channels. IEEE Trans. Commun., 43(2):773-85, Feb 1995. 4. R.R.H. Fischer and J.B. Huber. A new loading algorithm for discrete multitone transmission. In Proc. IEEE GLOBECOM'96,724-8, Nov 1996. 5. F. Tang, L. Deneire, M. Engels, and M. Moonen. Adaptive link adaptation. In Proc. IEEE GLOBECOM'01, Nov 2001. 6. J. Medbo and P. Schramm. Channel models for hiperlan/2 in different indoor scenarios. ETSIEP BRAN 3ER1085B, Ericsson Radio Systems AB, Mar 1998. 7. ETSI. ETSI ts 101 475 vl.1.1 broadband radio access networks (bran); HIPERLAN type 2; physical layer. Technical report, ETSI, Apr 2000. 8. Andrea J. Goldsmith and Pravin P. Varaiya. Capacity of fading channels with channel side information. IEEE Trans. Inform. Theory, 43(6): 1986-92, Nov 1997. 9. F. Tang, L. Deneire, M. Engels, and M. Moonen. A general optimal switching scheme for link adaptation. In Proc. IEEE VTC FALL'01, Oct 2001.

ON THE PERFORMANCE OF THE UBIQUITOUS ANTENNAS FOR THE RECEPTION OF COFDM SIGNALS SHUTAIOKAMURA AND SHOZO KOMAKI Department of Communications Engineering, Faculty of Engineering, Osaka University 2-1 Yamadaoka, Suita, Osaka, 565-0871, Japan E-mail: [email protected] MINORU OKADA Graduate School of Information Science, Nara Institute of Science and Technology, 8916-5 Takayama, Ikoma, Nara, 630-0101, Japan E-mail: mokada @ is.aist-nara.ac.jp This paper investigates the impact of a ubiquitous antenna system to the performance for the reception of coded orthogonal frequency division multiplexing (COFDM) signals. The ubiquitous antenna system is composed of multiple radio base stations (RBS) deployed over the service area, and each RBS is connected to the central control station (CCS) with radioon-fiber (RoF) link. In order to overcome the delay problem, we propose employing COFDM, which is robust to delay spread. The computer simulation results show the system can drastically improve the throughput performance compare with a centralized array antenna system.

1

Introduction

Recently, IEEE 802.11a wireless local area network (W-LAN) standard has been released [1]. IEEE 802.11a provides wireless local area network with data rate of up to 54Mbit/s at the occupied frequency bandwidth of 16.6MHz at 5GHz band. It employs orthogonal frequency division multiplexing (OFDM) technology in order to operate in heavy multipath environment. However, the bandwidth assigned for the system is not enough for deploying the radio base stations (RBS) over the service area. In Japan, the frequency band from 5.15GHz to 5.25GHz has been assigned for this 5GHz W-LAN system, and this band accommodates only four physical frequency channels in a 100MHz bandwidth. This makes the frequency allocation for each base station difficult when a lot of RBSs are deployed. On the other hand, we have already proposed ubiquitous antenna system, which is capable of canceling co-channel interference [2]. The ubiquitous antenna system is composed of multiple RBSs deployed over the service area and connected to the central control station (CCS) with radio-on-fiber (RoF) link. Since all the received signals at the RBS are collected to the CCS, we can employ more sophisticated signal processing algorithms such as interference cancellation and joint detection [3]

295

296

CCS E/O.O/E MODEM Signal Processing

Figure 1. The system description of the proposed ubiquitous antenna system

at the CCS. The proposed ubiquitous antenna system enables the multiple mobile terminals (MT) to operate at the same frequency simultaneously by making effective use of co-channel interference cancellation. Therefore the proposed system can improve the channel capacity. However, the propagation delay difference in RoF links makes it difficult to perform the interference cancellation and joint detection at the CCS. In [2], the propagation delay due to RoF link has not been taken into account. In this paper, we assume the COFDM system and consider applying the ubiquitous antenna system to the W-LAN in order to improve the channel capacity. Furthermore, we evaluate the performance of the ubiquitous antenna system by computer simulations. 2

2.1

System Model

Ubiquitous Antenna System

Figure 1 illustrates the system description of the proposed ubiquitous antenna system. All the RBSs are deployed over the service area in every hundreds meters. The RBSs are connected to the CCS with the RoF link. All the MTs in the service

297 area transmit the COFDM signals simultaneously at the same frequency. The transmitted signals are propagated through the multipath channels and received at the plural of the RBSs. The received radio frequency (RF) signal at each BS is applied to E/O converter to modulate the intensity of the optical carrier. The modulated optical signal is sent to the CCS via the RoF link. At the CCS all the optical signals from the RBSs are converted to RF signals again by the corresponding O/E converters. The CCS then demodulates the signals. Since all the MTs are transmitting the signals simultaneously at the same frequency, a multiple user detection algorithm is required for receiving all the signals transmitted by the MTs. In the following, we employ the minimum mean square error (MMSE) based algorithm as described in the next section [4]. Since the spacing amongst RBSs is around hundreds meters, about several hundreds nano-seconds of the delay difference due to RoF link. This difference in delay makes the multiple user detection difficult. In order to overcome this problem, we assume that the guard interval of the COFDM signals is longer than the delay difference. 2.2

Co-channel Interference Cancellation Based on Ubiquitous Antenna

The interference canceller considered in this paper is shown in Fig. 2. Suppose that M MTs transmit COFDM signals simultaneously at the same frequency channel. At each MT, the binary bit stream are coded into am[n,k]'s across the each COFDM subcarriers using convolution code, where n and k denote the symbol index and the subcarrier index, respectably. am[n,k]'s are then modulated into xm[n,k]'s using DPSK modulation. In order to estimate the channel impulse response (CTR) between MT and each RBS, the transmitter insert the pilot signals before the information data symbols. The pilot signal of each MT can be expressed as

xm[0,k]=Xl[0,k]W-KK°mk, where, WK=exp(-j2ii/K) is the complex Fourier kernel, and K0 denotes the largest integer smaller than or equal to K/M [5][6]. xm[n,k]'s are then applied to IFFT processor for multicarrier modulation with K subcarriers. The transmitted signal from each MT is received the L RBSs through the multipath channels whose frequency response is expressed as H

m = [^ml'"''^mt.] ' where x denotes the transpose of x. In this paper, we assume that Hml for different m's of Z's are independent. The received signal at the Z-th RBS is given by T

yi=ljHmlxm+zl, m=\

298 Input xjn,k] xm[n,t] am[n,k] Data Bit Conv. DQPSK GI # • IFFT Insertion Corder Modulation

rn

Pilot Insertion Transmitter MMSE-DC y,[n,k] xjn,k]

»FFT • »FFT

Normalized byMSE

yjn,k]

DQPSK Demodulation

€*• Soft-Dicision FEC Decordei

H,opt Channel Estimator

Output Data Bit — •

Pilot Generator Receiver Figure 2. Co-channel interference canceller based on ubiquitous antenna system

where zi denotes the additive white Gaussian noise, zero mean and variance o„2, at the f-th RBS. In order to simplify the notation, we define the received vector as

y=iyvyv--yJ

•

The receiver first estimates the CER by calculating the correlation between the received signal at each RBS and pilot signal. The estimated O R is expressed as

K, W = TT 2^xm [0, t]y, [0, t + r] . 0 1=0

And then, the receiver obtains Hml applying hmi[i] into the FFT processor. However, when signal-to-noise ratio (SNR) is small, the estimation error of the CER has been large, especially at the RBS far from the MT. Hence, the performance of the interference cancellation is degraded by using the received signal from such RBSs to interference cancellation. In order to overcome this degradation, the MMSE combiner performs the interference cancellation by using only L0 (smaller than or equal to L) RBS's signals, whose desired signal power is higher, and decides L0 depending on SNR. For systems with large SNR, since the estimation error of the

299

OR is small, Lo should be larger to obtain the diversity gain. However, for systems with small SNR, L0 should be smaller to avid the impact of the estimation error of the CTR.

10 20 Eb/NO (dB)

30

Figure 3. BER performance of the proposed MMSE-based interference canceller with MSE weighting

Before interference cancellation is performed, the received signals from the RBSs are divided into each subcarriers by the FFT processors. Then the MMSE based interference cancellation is performed in subcarrier-by-subcarrier basis. The weights of the MMSE combiner are given by -l *

where Ryy is correlation matrix of y defined as yH denotes the Hermitian transpose of y. The output of MMSE combiner for x0 is given by

^o=H0„y> The output of the MMSE is then applied to the FEC decoder. In order to improve the bit error rate (BER) performance after FEC, we employ soft decision decoding. The each output of the MMSE combiner is normalized by the corresponding mean square error (MSE) given by 2

0=CT d 2 -H o X H o>

where rjd is the desired signal power. The output signals normalized by MSE are then applied to soft-decision FEC decoder to decode the information bit stream. The

300 MSE normalization allows the FEC to get the subcarrier frequency diversity effect. Fig. 3 shows the BER performance of the proposed MMSE-based interference canceller with MSE weighting in an equal gain two-sample-spaced two-ray Rayleigh fading channel. We assume two MTs and two RBSs. As a transmission signaling, 48-carrier DQPSK with half-rate convolutional error correction coding is employed. Desired-to-undesired signal power ratio is OdB. The MSE normalization gives the power gain of lOdB at BER=10"5.

3

3.1

Simulation Results

Simulation Configurations RoF Link

\

CCS

* St ^

.^ J

h

\

7

^7

h ^ 100m (a)

JL_

Co-channel Terminals (b)

Figure 4. Simulation model

Now we analyze the performance of the proposed ubiquitous antenna based W-LAN system by computer simulation. In the following, we evaluate the up-link performance in synchronous symbol timing. In this case, all the COFDM signals from the MTs are synchronized in symbol timing at the RBSs. The configuration of the system in order to run the simulation is shown in Table 1. The allocation of the RBSs in order to cover the service area is shown in Fig. 4. We assume four RBSs throughout the analysis. In the proposed ubiquitous antenna system shown in Fig. 4(a), the service area is divided into four cells and the each RBS is located at the center of the cell. All the RBSs are connected to the CCS with the RoF link. All the MTs transmit the COFDM signal simultaneously at the same frequency channel. For comparison purposes, we also evaluate the performance of the adaptive array

301

antenna system whose antenna elements are located at the center of the service area as illustrated in Fig. 4(b). In the following, we referred to it as the conventional system. Again, all the MTs in the service area transmit the COFDM signal simultaneously at the same frequency channel. Table 1 Simulation Configurations

4

Data Length

10 symbols

4

Pilot Length

1 symbol

Uniformly Distribution

FEC

Convolution Constraint Length=6 Code Rate= 1/2 2-ray Rayleigh Fading

Number of MTs Number of BSs MT position Modulation FFTSize

7I/4DQPSK-OFDM 64

Number of Subcarriers

48

Channel

Bit Rate Bandwidth Symbol Rate Guard Interval Cell Size

11.5Mbps 20MHz

4.2ns

Radio Link Delay RoF Link Delay

500ns

Ills 100m

D/U

OdB

3.2

100ns

BER Performance of Proposed System

Firstly, we have evaluated the BER performance of the proposed ubiquitous antenna system. We assume that there is one MT in each cell in the ubiquitous antenna system, and the distance amongst a MT and the corresponding RBS is 35 meters. In order to evaluate the performance for the worst case scenario, all the MTs are located around the center of the service area. On the other hand, in the central antenna system, we assume that all the MTs are also located at 35 meters from the central RBS for the sake of fair comparison. The BER performance is shown in Fig. 5. The BER performance without interference cancellation (IC) is intolerable to transmit the data in both models. In contrast to the performance without IC, IC drastically improves the performance. The performance of the proposed ubiquitous antenna system is slightly better than that of the central antenna system in low Et/N0 region. However, mere is no major difference in the BER performance of both systems with IC.

302

10°

1 W^-— ~B^r

wifho IC without •

m Central antenna

10Pi W

Ubiquitous antenna

m

io -l

10 10 Eb/NO (dB) Figure 5. Bit error rate performance of proposed system

0.7

CO

OH

0.5

Limit

.

Ubiquitous a n t e n n a ^ - - - o - ~ / ^ with IC! o ~^ ^^ •

rh

Ubiquitous antenna/ w/oIC

3

SP 0.3

I


0.1

i

•

i

10 20 Nominal Eb/N0{$B)

Figure 6. Throughput performance of proposed system

•

30

303 3.3

Throughput Performance of the Proposed System

Secondly, we have evaluated the throughput performance of the proposed ubiquitous antenna system. In the following evaluation, we assume that the positions of all the MTs are uniformly distributed over the service area. The throughput per cell against the nominal E^N0 is shown in Fig. 6. The nominal Ej/N0 is E)/N0 when the distance amongst a MT and the nearest RBS is 70 meters. In case of central antenna system without IC, the throughput is 0.15 at the nominal E//N(j=30dB, while the proposed ubiquitous antenna system without IC gives three times higher throughput than the central one. IC further improves the throughput performance. The ubiquitous antenna system with IC gives better performance in low Ei/No region. Fig. 7 shows the throughput performance against the cell size at £yW^=30dB. The throughput performance becomes worse when the cell size exceeds 200 meters in diameter, since the delay spread due to the propagation delay difference amongst RoF links exceeds the guard interval of the OFDM signal. In order to employ the proposed OFDM based ubiquitous antenna system to larger cell size configuration, we have to manage this delay difference amongst the RoF links. Although we can also make use of longer guard interval in order to overcome the delay problem, the longer the guard interval is the worse the throughput. 0.7

Ubiquitous antenna with IC

\

Limit

0.61 t a M

0.5-| 0.4

g

• Ubiquitous antenna w/o

0.3 0.2 100

200

300

Cell Size (m) Figure 7. Throughput performance against the cell size at Et/Ng=30tiB

400

4

Conclusions

In conclusion, we have proposed a new ubiquitous antenna system, which employs COFDM, and investigate the performance of a ubiquitous antenna based W-LAN system for the reception of COFDM signals. From the results of the computer simulations, we have shown the proposed system can remove the delay due to the RoF link and improve the performance compare with conventional system. References 1. R. van Nee, G. Awater, M. Morikura, H. Takanashi, M. Webster, and K. W. Halford, New High-Rate Wireless LAN Standards, IEEE Commun. Mag., pp. 82-88, Dec. 1999. 2. M. Toyama, M. Okada, S. Komaki, Maximal Ratio Combining Macro Diversity for Micro-Cellular Slotted ALOHA, IEICE B-I, vol.J79-B-I, 5 pp. 271-277, May 1996. 3. S.J.Grant, J.K.Cavers, Performance Enhancement Through Joint Detection of Cochannel Signals Using Diversity Arrays, IEEE Trans. Comm., vol.46, No.8, pp. 1038-1049 Aug. 1998. 4. Y. Li, R.Sollenberger, Adaptive Antenna Arrays for OFDM Systems With Cochannel Interference, IEEE Trans. Comm., vol.47, No.2, pp. 217-229 Feb. 1999. 5. Y. Li, N. Seshadri, and S. Ariyavisitakul, Channel Estimation for OFDM Systems with Transmitter Diversity in Mobile Wireless Channels, IEEE Journal on Selected Area in Communications, vol.17, pp. 461-471, Mar. 1999. 6. M. Miimster, L. Hanzo, Improved Decision-Directed Channel Estimation for Multi-User OFDM Environments, in Proc. of Vehicular Technology Conference, (Rhodes, Greece), IEEE, May 2001. 7. M.V.Clark, T.M.Willis, L.J.Greenstein, AJ.Rustako.Jr., V.Erceg, R.S.Roman, Distributed Versus Centralized Arrays in Broadband Wireless Networks, in Proc. of Vehicular Technology Conference, (Rhodes, Greece), MA1-2, IEEE, May 2001.

IMPROVED AUTOMATIC FREQUENCY CONTROL FOR OFDM HOSSEIN HOSSEINI AND BIJAN ROHANI

Genista Research Pte. Ltd. 1 Kim Seng Promenade #14-04 Great world city East Tower Singapore 237994 E-mail: [email protected] E-mail: brohani @aenista.com.sa Frequency eiror can result is rapid BER degradation in OFDM systems. Preamble symbols are commonly employed for estimation of the frequency offset. Pilot symbols can be used for improving the frequency error estimation obtained from preamble symbols. The frequency offset for detection of each OFDM symbol can be estimated from the pilot symbols in the most recent OFDM symbol as well as previous symbols. In this paper, the use of pilot symbols for frequency error estimation and compensation is considered. A frequency estimation algorithm with memory is proposed wherein some portion of the pilots from previous OFDM symbols are used. It is shown that depending on memory of the algorithm, the sensitivity of the OFDM BER can be changed. Results show that the BER performance is least sensitive to frequency offset when the memory of algorithm is zero, i.e. when the pilot symbols from previous OFDM symbols are not considered. However, in this case the BER is lager than it would be when the memory is non-zero. It is shown in this paper that the best compromise between the BER and frequency offset sensitivity can be achieved when 20% of the strength of pilot symbols from previous OFDM symbols is used.

1

Introduction

The Orthogonal Frequency Division Multiplexing (OFDM) has been adopted as the access technique of choice for a variety of applications including digital audio broadcasting and television broadcasting systems [1], [2], and Wireless Local Area Networks (WLAN) [3]. These applications essentially require high quality transmission at high data transmission rates. In wireless local area networks, where data transmission at high rates is hampered in the presence of frequency-selective fading, OFDM is an effective technique [4], [5]. However, the performance of OFDM could be adversely affected as a result of frequency synchronization errors [5], [6]. In this paper the use of pilot symbols for improving frequency offset estimation is proposed. In Sections 2 and 3, packet synchronization and frequency acquisition for OFDM packets are described. Imperfect packet synchronization and frequency acquisition give rise to a phase error, which influences the BER performance of the receiver. The pilot-aided frequency compensation scheme, which is proposed for reducing the phase error, is described in Section 4. This is followed by the

305

306

presentation and discussion of some simulation results in Sections 5 and 6, respectively. 2

Timing Synchronization

In OFDM, preamble symbols are transmitted in each packet for timing estimation for packet synchronization and subsequent compensation of frequency errors. This preamble contains a sequence of ten identical short symbols of 0.8|xs duration each followed by two long symbols of 4jxs duration each. The short symbols are employed for timing synchronization while the long symbols are used for frequency compensation. For packet synchronization the start of the packet has to be detected as accurately as possible. Packet synchronization is done in two stages, namely; coarse synchronization and fine synchronization. In coarse synchronization the crosscorrelation between the known short symbols and the received symbols is calculated. Once the cross-correlation exceeds a specified threshold, the crosscorrelation is observed over a short interval thereafter. The cross-correlation exhibits ten consecutive peaks corresponding to the short symbols. These peaks are of approximately the same strength and are evenly spaced as shown in Figure la&b. The fine synchronization is achieved by adjusting the timing based on crosscorrelation peaks. The accuracy of the fine synchronization depends on the sharpness of crosscorrelation peaks. Sharp peaks, such as those shown in Figure la, result in very accurate timing. However, the cross-correlation peaks are generally not sharp. An example of such cross-correlation peaks is shown in Figure lb. These broad peaks affect the accuracy of the timing estimation. As a result the timing error At is increased. This timing error is translated to a phase error Acp proportional to Af. The phase error Acp affects the OFDM BER performance therefore it is necessary to reduce it before symbol detection. This error can be reduced as part of frequency acquisition process.

307 Match Filter For Short Symbols

Match Filter For Short Symbols

TimG in/is (b) Figure 1 the cross-correlation between received data and actual short symbols in the receiver (a) sharp peaks, (b) broad peaks.

308

3

Frequency Acquisition

After time synchronization, it is necessary to do frequency acquisition. Two long symbols are inserted for this purpose in each packet. The frequency error can be estimated based on these two symbols as shown in Figure 2. As shown in Figure 2, the received long symbols are transformed into frequency domain by calculating their respective FFTs. The delay element is inserted in the signal path so that the timing of the first and the second long symbols is aligned. After the FFT of the two symbols are computed, they their complex-valued product is calculated. The phase of the resulting product is extracted by the function arg(.). The phase is then averaged over the duration of one symbol. This average corresponds to the estimated frequency error or frequency offset. The frequency error is reduced by removing the estimated frequency error from the OFDM symbols before symbol detection.

Signal

FFT Noise Arg(.) Delay

Averager

Frequency "•

Offset

FFT

Figure 2 Block diagram showing the principal steps in estimation of phase error.

The effectiveness of frequency offset compensation depends on the accuracy of the estimated offset. There will still be some residual frequency error even after offset compensation. The Probability Distribution Function (PDF) for the residual frequency error has been shown in Figure 3. This PDF has been obtained for frequency error based on simulation of OFDM physical layer for 15000 transmitted packets. The transmission channel is based on IEEE 802.11a model with 18 multipath components [7] with Signal-to-Noise ratio set to 10 dB. It should be noted that the absolute values of frequency offset have been used in plotting Figure 3. As observed from Figure 3, the error PDF follows the Gaussian distribution. The standard deviation and mean of frequency error are approximately 3 KHz and 1 kHz, respectively.

309 Distribution of error in frequency offset detection

o

2000

1000

6000

Frequency Offset Figure 3: Distribution of error infrequency-offsetdetection 4

Pilot-Aided Frequency Compensation

In WLAN standards there has been provision for pilot symbols that are intended for updating the estimated channel information during each packet. These pilot symbols are inserted in each OFDM symbol. The pilot symbols can also be useful in improving the accuracy of frequency error estimation that have akeady been calculated from preamble symbols. However, to the authors' knowledge, pilot symbols have not been used for this purpose yet [8-10]. In this Section, the use of pilot symbols for frequency error estimation and compensation is considered.

The frequency offset for detection of each OFDM symbol can be estimated from the pilot symbols in that same OFDM symbol. The frequency error is estimated from the average value of the pilot symbols. This average value is subtracted from every subcarrier in the OFDM symbol so that the frequency offset is compensated for. The

310

accuracy of the estimation can be further improved by using the pilot symbols from previous ODFM symbols. The pilots from previous OFDM symbol can be used for calculating the offset in the current OFDM symbol according to,

/.(#0 = ( l - a ) / » + a/.(n-D

(D

where, fe(n) is the estimated frequency error in the nth OFDM symbol and a is constant; 0< a<\. In addition, fe(n) is the average error in the received pilots for the nth OFDM symbol, which is calculated according to,

l(n) = {l/4)f,[p(i,n)-p(i,n)}

(2)

i=l

where, p(i,n) is the ith received pilot in the nth OFDM symbol and p(i,n) is the corresponding transmitted pilot. 5

Simulation Results

The performance of automatic frequency control scheme described above has been investigated by computer simulation. In this simulation, IEEE 802.11a physical layer [7] has been used. The packet size used was 64 bytes at a data rate of 12 Mb/s. In this simulation, the bit energy per noise spectral density, Eb/N0, was set at constant value of 12 dB. With Eb/N0 fixed at this value, the average Bit Error Rate (BER) of the received data has been obtained as function of carrier frequency offset. The results of simulation for several values of the parameter a are shown in Figure 4. From Figure 4, a= 0.2 is the best choice for the memory parameter. This value gives the best compromise between BER and sensitivity to frequency offset. The BER performance of the proposed scheme with a=0.2 in comparison with the conventional method, i.e. a=0, is shown in Figure 5. In addition, more simulations were carried out to investigate the sensitivity of the proposed scheme at different transmission rate with a=0.2. These results are shown in Figure 6.

311

-•"without AFC -»-VMthAFC(a=0.5) "*"with AFC (a=0.8) -g-withAFC(a=0.2) - » " with AFC (a=0) 00001

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

10000

Frequency Offset (Hz)

Figure 4: BER performance of OFDM system as a function of frequency offset.

I

-»-a=0 -*-a=0.2

L

-~~-^—

Eb/No (dB)

Figure 5: Comparison of BER performance of the proposed scheme with conventional method at transmission rate of 12 Mb/s.

312

-*-Rate=12 Mb/s Mb/s

: :

-*-Rate=24Mb/s

•

I l_

0

1000

2000

3000

4000

-

-

-

-

•

—

5000

6000

7000

8000

Frequency offset Figure 6: Sensitivity of the proposed scheme (a=0.2) to frequency offset at various transmission rates.

6

Discussion

As it is observed from Figure 4, the BER performance is very sensitive to frequency offset when no AFC is employed. The BER is degraded by more than one order magnitude if the frequency offset is increased to 2000 Hz. The performance is improved by employing AFC. When a=0; i.e. only pilot symbols from the current OFDM symbol are used for AFC, the performance is least sensitive to frequency offset. However, in this case, the BER is higher than for non-zero values of a. The sensitivity to frequency offset is increased with the value of a, but lower BER values are achieved as a result. The best trade off between BER performance and frequency offset is achieved for a=0.2. The performance curve for a=0.2 and a=0 cross over at 6000 Hz frequency offset. However, according to the PDF of frequency offset, shown in Figure 3, the offset almost always is less than 6000 Hz. Additional simulation results presented in Figure 5 show that there is approximately 1 dB gain in the performance of the proposed scheme relative to the conventional method at Eb/N0=12 dB and data rate of 12 Mb/s. Similar gains can be expected for other data rates. Results shown in Figure 6 highlight the sensitivity of the proposed scheme with a=0.2 to frequency error at several transmission rates. These results show that, in general the sensitivity of this scheme to frequency offset is low. In addition,

313 sensitivity does not necessarily increase or decrease with data rate as a general rule. This is because sensitivity to frequency offset is not only a function of frequency error alone, but difference in modulation and channel coding applied at different rates also affect the sensitivity.

7

Acknowledgements

We thank Riaz Esmailzadeh and Anand R. Prasad for reviewing the paper and helping us improve it with dieir comments. References 1. K. D. Kammeyer, U. Tuisel, H. Schulze, and H. Bochmann, "Digital multicarrier-transmission of audio signals over mobile radio channels," European Trans. Telec, vol. 3, no. 3, May-June 1992. 2. H. Sari, G. Karam, and I. Jeancaude, "Transmission techniques for dig-ital terrestrial TV broadcasting," IEEE Commun. Mag., Feb. 1995. 3. G. A. Halls, "HIPERLAN: The high performance radio local area net-work standard," TEE Elec. Commun. Eng. J., vol. 6, Dec. 1994. 4. L. J. Cimini, "Analysis and simulation of a digital mobile channel using orthogonal frequency division multiplexing," IEEE Trans. Commun., vol. 33, pp. 665-675, July 1985. 5. M. Speth, D. Daecke, and H. Meyr, "Minimum overhead burst synchronization for OFDM based broadband transmission," in Proc. Globecom'98, pp. 32273232. 6. M. Speth, F. Classen, and H. Meyr, "Frame synchronization of OFDM systems in frequency selective fading channels," in Proc. VTC'97, pp. 1807-1811. 7. Draft Supplement to Standard for Telecommunications and Information Exchange Between Systems - LAN/MAN Specific Requirements - Part 11: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications: High Speed Physical Layer in the 5 GHz band, IEEE P802.ll Wireless LANs d o c : IEEE 802.11-98/370ar3, Nov. 1998. 8. Blind frequency synchronisation for OFDM systems with pulse shaping Manimohan, V.B.; Fitzgerald, W.J. Communications, 1999. ICC '99. 1999 IEEE International Conference on, 1999 Page(s): 178 -181 vol.1

314

9. Decision-directed maximum-likelihood estimation of OFDM frame synchronisation offset Keon Woo Kang; Jaemin Ann; Hwang Soo Lee Electronics Letters, Volume: 30 Issue: 25, 8 Dec.1994 Page(s): 2153 -2154 10. Low-complex frame synchronization in OFDM systems, van de Beek, J.-J.; Sandell, M.; Isaksson, M.; Ola Borjesson, P. Universal Personal Communications. 1995. Record., 1995 Fourth IEEE International Conference on , 1995 Page(s): 982 -986

Part 11 Antennas and Interference Control

A MEVIO ARCHITECTURE FOR WIRELESS INDOOR APPLICATIONS LUCA GIANGASPERO, GIOVANNI PALTENGHI CEFRIEL, via Fucini 2, 20133 Milano (Italy) E-mail: [email protected], Tel: +39-2-23954338, Fax: +39-2-23954254 LUIGIAGAROSSI Philips Research Monza, via Casati 23, 20052 Monza Ml (Italy) E-mail: [email protected], Tel: +39-39-2037802, Fax: +39-39-2037800 This paper deals with the performance evaluation of a Multiple Input Multiple Output (MIMO) system in ideal and non-ideal propagation conditions. The system operates in a wireless indoor environment in the 17 GHz frequency band. Space Time Coding and Space Division Multiplexing techniques have been considered. A multicarrier (OFDM) MIMO architecture that significantly increases the achievable bit rate of the system has been studied and analyzed. Simulation results show the effectiveness of the considered architecture.

1

Introduction

The development of wireless communication systems for high bit rate data transmission and high-quality information exchange between terminals in an indoor environment is becoming one of the new challenging subjects in the telecommunication research area. The market demand for broadband multimedia services, ubiquitous networking and Internet access via portable devices is expected to grow enormously pushing the development of modem architectures for high bit rate transmission. The European project WIND-FLEX [12], that is the framework of this paper, is certainly a pioneer in this new context. Aim of the project is to design and develop a high bit rate flexible and configurable modem architecture, which works in singlehop ad hoc networks and which provides a wireless access to the Internet for slowly moving users (about 1 m/s) in an indoor environment. The scenario is then a typical home environment with a link-range set to 20 meters LOS (Line Of Sight). A basic requirement for the WIND-FLEX modem is the capability of managing variable bit rates ranging adaptively from 64 kbit/s to 100 Mbit/s of payload, to bring new services and to trigger new interesting and appealing applications with different required QoS (Quality of Service). The 17 GHz unlicensed frequency band has been chosen since the frequency bands around 2.4 GHz and 5 GHz have already been allocated and are close to saturation. Moreover, the characteristics of the 17 GHz transmission channel sound rather suitable for the WIND-FLEX system and the choice seems to be a promising and challenging solution. The available band of 200 MHz (17.1-17.3 GHz) is divided into four 50 MHz-width channels not simultaneously selectable. A

317

318 multicarrier modulation scheme (OFDM, Orthogonal Frequency Division Multiplexing) is designed for an efficient wireless link in one of these 50 MHz wide channels. Although the capacity of the WIND-FLEX system is already rather high with respect to the other existing wireless systems, it sounds worth to work out the idea of a further increase exploiting a MTMO (Multiple Input Multiple Output) system architecture. The basic idea is to usefully exploit the multipath, rather than mitigate it, considering the multipath itself as a source of diversity that allows the parallel transmission of a certain number of sub-streams from the same user. This approach is based on a MTMO (N,M)-system (N transmitting and M receiving antennas) which supports transmit space diversity obtaining higher data rate while using the same total power and bandwidth adopted for the 1-D (one transmitting and one receiving antenna) system. The indoor environment is the ideal rich scattering environment necessary to get the best performance promised by this approach. Space Time Coding (STC) and Space Division Multiplexing (SDM) algorithms exploit the rich scattering wireless channel allowing the receiver to detect the different signals simultaneously transmitted by the iV transmit antennas. STC permits to efficiently introduce a spatio-temporal correlation among transmitted signals to improve information protection while SDM goal is to increase data rate. Several different Space Time Coding techniques have been considered: Space Time Trellis Codes [1,2], Space Time Block Codes [3], D-BLAST [4]. All of them appear to be too complex, especially at the receiver side, for our specific application and so Space Division Multiplexing [5] sounded worth to be investigated instead. Among SDM techniques, V-BLAST [6,7] seems to exhibit the best trade-off between performance and complexity. V-BLAST algorithm implements a non-linear detection technique based on Zero Forcing (ZF) combined with symbol cancellation to improve the performance. The idea is to look at the signals from all the receive antennas simultaneously, first extracting the strongest sub-stream from the received signals, then proceeding with the remaining weaker signals, which are easier to recover once the strongest signals have been removed as a source of interference. This is somewhat analogous to the decision feedback equalization. When symbol cancellation is used, the order in which the sub-streams are detected becomes important for the overall performance of the system. In fact, the transmitted symbol with the smallest post detection SNR {Signal to Noise Ratio) will dominate the error performance of the system. Post detection SNR is determined by ordering. The optimal ordering is based on the result that simply choosing the best post detection SNR at each stage of the detection process leads to the maximization of the worst SNR over all possible orderings [6].

2

Analyzed Architecture

The analyzed architecture represents a natural extension of the multidimensional signal transmission involving space dimension. Transmit space diversity techniques

319 require flat fading in the channel bandwidth. The use of OFDM allows considering this hypothesis true in each sub-band [8]. Therefore an evolution of the V-BLAST supporting OFDM modulation seems to be a promising architecture in order to match the requirements of modern wireless modem. 2.1

Transmitter

The transmitter is equipped with N antennas (see Figure 1). A traditional 1-D channel encoder codes the source information bits. The code bits are then mapped on the symbols of the constellation adopted for each frequency channel, which corresponds to an OFDM subcarrier (in WIND-FLEX these are M-QAM symbols). The vectorial nature of the transmission is introduced by demultiplexing (1—>N) these symbols. In the considered architecture this demultiplexer represents the space encoder. It maps symbols on the N space channels, which are sub-streams of the same user. A serial to parallel converter for each space channel takes L of these symbols to form the input for the OFDM modulator. L is the number of frequency channels, that is the number of the OFDM subcarries (in WIND-FLEX L=128). The corresponding antenna transmits the output from each modulator. This structure permits the simultaneous transmission of N-L M-QAM symbols in the same bandwidth and with the same total power required by an OFDM symbol for carrying only L of them. lomnTX/

y^ y^ V^

y^ Figure 1. Transmitter architecture.

320

2.2

Receiver

The receiver is equipped with M antennas; note that V-BLAST requires M>N. Each antenna receives a different noisy superimposition of the faded versions of the N transmitted signals (Figure 2). If the transmit and receive antennas are sufficiently spatially separated, more than X/2 (at 17 GHz it is about 0.9 cm) and there is a sufficiently rich scattering propagation environment, the transmitted signals arriving at different receive antennas undergo uncorrelated fading. Moreover, if the channel state information at the receiver is known, that is, if the channel matrices Hi (one for each frequency channel) are known, it is possible to detect the N space channels. The output of the OFDM demodulator, at the receive antenna m, is a set of L signals, one for each frequency channel, described by N

r

mi

m,l

= Y Hm „ i • C„, + T]m,, with I = 1,..., L, / i n=l

m,n,i

n,l

im,l '

'

'

'

where fl„i is the flat fading coefficient representing the channel from the transmit antenna n to the receive antenna m at frequency /, and t]mj are independent samples of a Gaussian random variable with power spectral density N0 representing noise (note that A^ is the power spectral density of the noise at the receiver input [5]). The M outputs for die frequency / are the inputs to a V-BLAST signal processor. This sub-system is able to detect the A' different space channels once flat fading is assumed. OFDM modulation-demodulation approach makes this hypothesis, for each frequency channel, reasonable. This processing is repeated for each of the L sub-bands. The N-L M-QAM symbols, obtained at the output of the L V-BLAST signal processors, are then serialized by a parallel to serial converter, in order to complete the traditional 1-D receiver processing (Figure 3).

^-y

* —

K

J<

• pp, 1

ru = Z H3_,j-c.j +nv

--".Z.

-sS= i

*v
.__

»-

p.

,J=L •

"

Figure 2. Receiver architecture: frequency channels selection and space channels detection.

321 —

*

—

>•

_ ^ „:::;Hf:: — — •

-=—»•

r,.,.../,"*....i i

JJ. 1

J

*"

M

L

»* •

Figure 3. Receiver architecture: 1-D processing.

3

Channel Model

The SISO (Single Input Single Output) channel model adopted to calculate the Hmni coefficients is the one presented in [13]. The NLOS (Non Line Of Sight) channel is modeled with eighteen taps. The taps amplitude has been shaped following different probability density functions: a combination of exponential and WeibuU pdf s for the first bin and exponential pdf s for the others. In the considered context however, the SISO model is not sufficient to completely characterize the channel because the MIMO approach entails also the consideration of the spatial correlation characteristics of the transmission system. 3.1

Spatial Correlation

Several studies have recently demonstrated that if channel path gains of a (iV,M) MIMO system are completely uncorrelated, the channel capacity scales linearly with n for sufficiently high n, where n = mm(N,M). This improvement can be obtained even if the transmitter does not have channel state information [9]. The problem is that, in real-world propagation conditions, these channel coefficients could be partially correlated. Generally the correlation coefficients depend on many factors like the physical parameters of the transmit and receive antennas, i.e. antenna spacing, but also on the characteristics and the distribution of the scatterers. It has been observed that when fades are correlated, the channel capacity can be significantly smaller than when they are not. This suggests further research about spatial correlation problem [10,11]. Here, for the overall system

322

performance evaluation, the correlation has been taken into account by means of a synthetic parameter, that is the mean correlation coefficient of the MIMO channel. Some correlation matrices R (the generic element rtj of the correlation matrix is the correlation coefficient between the i-th and the j-th path gains) have been defined with an increasing value of the mean correlation coefficient ranging from 0 (totally uncorrected path gains) to 0.8 (almost completely correlated path gains) with steps granularity of 0.2. 4

Simulation Results

To verify the performance of the considered architecture in ideal and non-ideal conditions (i.e., the presence of spatial correlation between path gains), a simulator has been implemented. The BER (Bit Error Rate) as a function of the SNR (E^No) has been evaluated. 4.1

Ideal Conditions

The first set of performance curves has been found in ideal propagation conditions (i.e., the path gains can be considered uncorrelated). The dependence of the performance from N and M has been analyzed. Given a particular modulation, 16-QAM for instance, a system having the same number of transmit and receive antennas has been simulated. In Figure 4 curves for different N are reported. „ 10

(N,N) System-16-GAM -e-

_ . _ - - - - - - t?";-N>

- - - - - - -

\ \\N<x \ 1

' ^^V^V

- - - - - - -

'

• • 10

•-•*•- N = M = 4

.

" * : : ^o^~

5

N=M=2

- r - N=M=3

15 20 SNR at each RX antenna

I

^^sj. - - - -

'

'

25

^^s.

^S 30

Figure 4. System performance comparison for different values of N = M.

This figure confirms that the V-BLAST technique allows increasing the bit rate without having a significant BER worsening. Note that the total transmitted power

323

and the bandwidth are the same for all the (N,M) systems. This result has been verified for each modulation alphabets. In Figure 5 the performance curves for different M, using a 16-QAM modulation on each subcarriers, are shown. (N,M I System - 16-QAM

10°

: = = = I = = = E: 10"' ^^r^~—*~^jr~i

10*

! I ! I\

Eio J

^-^i_^»

:

=

z

z

:

- e - N=2M=2 1 - t - N=2M=3

- » - N=2 M=4 . *=--L= = - : - \ \ \ \ \ \ \

i fKK T T T H ^ , . „ -*

, r ^ A v = =!=- = =

MM MM

\

•4

;

z

1 = = 1 !

!

= 1

-v^

E : : : ! 1

= 1, 1 ? v !

: : ; :

5

r-k::

: : : : ; : :V : ; • : : :"s

10*

I

1 I

1 1 i

i

. . . . i A . .•. , . „

r

: : : ! 1 z -\ L

.n- 8

:' : : ! :

\J

SNR at each RX antenna

Figure 5. System performance comparison for different values of M.

It can be easily seen that increasing M means improving the performance. This is due to the availabibty of a greater number of received signals that can be combined in a more efficient way to obtain more accurate estimate of the transmitted signals. Some other simulations indicate that the system performance improvement, compared to the N = M case, is function of the difference M - N, while resulting substantially independent from N. Figure 6 underlines this property. (N,M) System - 16-QAM

in-

i M r i i =r i i =i i i i i r i i i i to"*

10*

E io J m

p.

11 u 1 ?> feli^v! I n H U i f M M' I I M 1 Mx= ! Fix! M i l l . - . . . ,

10"*

•6

10

m-

- . - , - - .

= = = ='i =i -S-t* -«*5

N=M=2 N-M=3 N=2M=3 N=3M=4 N=2M=4 N=3M.S

,

-x^

-

- v j v v -i -

-

-

ill i i i h i i l H i r

: : : : : : :

c i \ : : : : :>>i

,

Y

,

^

5 M I M I I M \ | I Si I I ,

,

15 20 SNR at each RX antenna

is 25

Figure 6. System performance comparison for different values ofM-N.

324

4.2

Non-Ideal Propagation Conditions

The non-ideal propagation condition analysis has been limited to the (2,2) and (3,3) MEMO systems. Figure 7 shows the simulations results for the (3,3) system with 4/16/64-QAM modulations on the subcarriers, for different values of the mean correlation coefficient among paths. The mean correlation coefficient ranges from 0 to 0.8 with steps of 0.2. 0

10 L -

.

lO"1!

5

.

_i -

1

10

-

.

(3,3) System-4-QAM _ i_ - . . L -

i

1

15 20 SNR at each RX antenna

„ 10 L . . .

(3,3) System - 16-QAM .1 . . . . I . . . . I .

10"3' 5

' 10

' ' 15 20 SNR at each RX antenna

1

25

30

' 25

30

325 (3,3) System -64-QAM 10° b^™zuZZr^..-

— ^ i

: : : •: :

2^r--Aj " "=~-^.-.;.

r ^ ""^^^

10''

— , — io" z

- *- -

: = = : : :E -ei .

5

.

*

•

e :

::•:::: - --*---- - • « - . : .

-

-

-

*£NJ:

5

: \ : =*

uncorrelated paths i

•

mean corr coeff.=0.4 mean corr coeff.=0.6

..^.. mean corr coefl. =0.8 , , 10 15

20 SNR at each RX antenna

,1 25

30

Figure 7. Performance of the (3,3) MIMO system for different mean correlation coefficients and modulations: 4-QAM, 16-QAM and 64-QAM, respectively.

The analysis and comparison of these curves show a non-linear relation between the performance degradation and the mean correlation coefficient. Moreover, the negative effect of the spatial correlation becomes more and more evident as SNR increases. Under a threshold SNR, the BER performance is limited mainly by the presence of the noise. The curves show that the SNR loss of the (3,3) MTMO system due to the spatial correlation does not exceed 5 dB up to a mean correlation coefficient equal to 0.6, in the considered SNR range. It must be pointed out that 0.6 is a high value of correlation. Beyond this value the use of a MTMO system would be inefficient for the specific propagation environment. The results for the (2,2) system exhibit similar properties. 5

Conclusions

In this paper multiple input multiple output transmission system based on the VBLAST Space Division Multiplexing technique and an OFDM modulation scheme has been proposed and evaluated in ideal and non-ideal propagation conditions. The simulation results show the capability of the analyzed architecture to greatly improve the bit rate without increasing the total transmitted power or the required bandwidth. The effectiveness of the analyzed architecture has been demonstrated also in case of non-ideal propagation conditions, i.e. correlation among path gains. The performance degradation is in this case Umited (< 5 dB) for a mean correlation coefficient < 0.6.

326 REFERENCES 1. A. F. Naguib, N. Seshadri, and A. R. Calderbank, "Increasing Data Rate over Wireless Channels," IEEE Signal Processing, pp. 77-92, May 2000. 2. V. Tarokh, N. Sehadri, and A. R. Calderbank, "Space-Time Codes for High Data Rate Wireless Communications: Performance Criterion and Code Construction," IEEE Transactions on Information Theory, vol. 44, pp. 744765, March 1998. 3. S. M. Alamouti, "A Simple Transmit Diversity Technique for Wireless Communications," IEEE Journal on Select Areas in Communications, vol. 16, no. 8, pp. 1451-1458, October 1998. 4. G. J. Foschini, "Layered Space-Time Architecture for Wireless Communication in a Fading Environment When Using Multi-Element Antennas," Bell Labs Technical Journal, Autumn 1996. 5. A. van Zelst, "Space Division Multiplexing Algorithms," 10th Mediterranean Electrotechnical Conference, MELECON 2000, vol. 3, pp. 1218-1221. 6. P. W. Wolniansky, G. J. Foschini, G. D. Golden, and R. A. Valenzuela, "VBLAST: An Architecture for Realizing Very High Data Rates Over the RichScattering Wireless Channel," invited paper, Proc. ISSSE-98, Pisa, Italy September 1998. 7. G. J. Foschini, G. D. Golden, R. A. Valenzuela, and P. W. Wolniansky, "Simplified processing for high spectral efficiency wireless communication employing multi-element arrays," IEEE Journal on Selected Areas in Communications, vol. 17, no. 11, pp. 1841-1852, November 1999. 8. D. Agrawal, V. Tarokh, A. Naguib, and N. Seshadri, "Space-Time Coded OFDM for High Data-Rate Wireless Communications Over Wideband Channels," IEEE VTC '98. 9. G. J. Foschini, and M. J. Gans, "On Limits of Wireless Communications in a Fading Environment when Using Multiple Antennas," Wireless Personal Communications, vol. 6, no.3, pp. 311-335, March 1998. 10. Da-Shan Shiu, G. J. Foschini, M. J. Gans, and J. M. Kahn, "Fading Correlation and Its Effects on the Capacity of Multielement Antenna Systems," IEEE Transactions on Communications, vol. 48, no. 3, pp. 502-513, March 2000. 11. J. W. Wallace, and M. A. Jensen, "Characteristics of measured 4x4 and 10x10 MIMO wireless channel data at 2.4 GHz," IEEE Antennas and Propagation Symposium, January 2001. 12. 1ST-1999-10025, WIND-FLEX D2.5, "The final modem specification for baseband, RF/IF block and upper layers." 13. M. Lobeira, A. G. Armada, R. Torres, and J. L. Garcia, "Parameter estimation and indoor channel modelling at 17 GHz for OFDM-based broadband WLAN," 1ST Mobile Comm. Summit 2000, pp. 29-33, Gal way, Ireland, October 2000.

INTER-CELL I N T E R F E R E N C E EFFECT O N O F D M - B A S E D WIRELESS L A N PETRI MAHONEN, ANTONY JAMIN University of Oulu, Finland Center for Wireless Communication Emails: [email protected], [email protected] The fast deployment of the unlicensed wireless local area network (WLAN) is raising some issues on the re-use of radio-frequency spectrum. Since any channel of an ISM band can be freely selected, it appears likely that the number of allocated channels is not sufficient to guarantee inter-cell disturbance free communication. This letter reports on the performance degradation of an OFDM link subject to disturbance by similar devices belonging to an adjacent network and using the same radio frequency channel.

1

Introduction

Recent developments in wireless communication have led to a growing interest for broadband WLAN networks. The first generation, largely based on spread spectrum modulation in order to deal with indoor propagation *, is being replaced by OFDM because of spectral efficiency. This modulation scheme has been successful in other application such as terrestrial broadcasting 4 . It has been seen as a suitable modulation scheme to provide the high capacity required within the limited available bandwidth. Wireless networks have been implemented successfully using OFDM e.g. the IEEE802.11a and Hiperlan 2 standards 2 ' 5 . However, this change of modulation scheme might lead to performance degradation when the number of available channel is not sufficient for spectral separation i.e. forcing adjacent cells to simultaneously make use of the same frequency carrier. Such a scenario is likely to occur for instance in a large building hosting number of independent small companies, each of them managing its own wireless network independently. In this paper, we therefore focus on analyzing the bit error rate (BER) degradation of an OFDM link in the presence of another similar OFDM disturbing signal. Our system model and the general assumptions made will first be presented. Parameters characterizing the transmitted signal of interest,the disturbing signal, and the channel model are then defined. Their actual influence is reviewed in the following section. Based on the results obtained, the relevance of using OFDM modulation in dense high speed cellular networks is finally discussed. 327

328 2

System model

We intend to evaluate the degradation of performance of an OFDM link when the received signal of interest is mixed with another self-similar OFDM signal transmitted by a closely located transmitter. The signal at the receiver side is then the sum of the desired signal, a second disturbing OFDM signal at the same carrier frequency, and channel noise. For simplicity, we are only assuming an AWGN channel so that no other distortion of the transmitted signal due to the channel is considered. With this system model, there are number of parameters that can be chosen independently for the main communication link and disturbing transmitter. We assume here that both sources are transmitting an OFDM modulated signal continuously. Though this is not an accurate model for typical frame-based WLAN, this represents an extreme case permitting usto study the interference effect on the modulation exclusively. The specific characteristics of this particular OFDM signal are taken from the IEEE802.11a standard 2 . However, no forward error correction or interleaving is assumed since we are interested in the effects on the lower physical layers of the system. Finally, perfect synchronization is assumed at the receiver since an analysis of the behavior of synchronization algorithms is out of the scope of this article. Nevertheless, this would definitively be a very valuable contribution on this topic. The most relevant parameters in this study are the ones characterizing the disturbing signal. In addition, the following parameters have to be taken into consideration: • A W G N channel noise power: this will be expressed as the equivalent SNR of the received signal (i.e. with disturber taken into account), • Disturber relative power: The disturbing signal power Pdist relatively to the signal of interest, • Disturber constellation: the constellation used by the disturbing signal can be any of the following: BPSK, QPSK, 16QAM, or 64QAM. • Time offset: the time delay At between the start of OFDM symbols of disturber and desired signals, expressed in number of sample periods Ts, • Carrier frequency offset: the carrier frequency offset AF between transmitters. The interaction of those parameters is shown with the system model in Figure 1.

329

^^e

OFDM transmitter

OFDM transmitter (disturber)

f

»H>-<S>

Dafay

/

At

AF i D i sturber

OFDM receiver

model

Usy-@ r

di«

AWGN

Channel model

Figure 1. Simulation model block diagram

3

Simulations setup

With the system modelled as denned in theprevious section, it is rather difficult to clearly identify the effect of each system parameter individually. Thus, our analysis is divided into four scenarios. In each case, the effect of each of the system parameters previously listed is analyzed. Other parameters are arbitrary chosen to their nominal (ideal) value. Obviously, none of those four scenarios is likely to be representative of the behavior of an actual implemented system since those system parameters can take any arbitrary value concurrently. Nevertheless, this approach permits a better understanding of the phenomena observed and permits a good approximation of the overall disturbance created. 3.1

Synchronous disturber with identical constellation

We assume for this scenario that the OFDM link under study is disturbed by another OFDM signal using an identical constellation. It is further assumed that the signals are totally synchronous. The effect of time offset between signal being studied in depth below. Figure 2(b) to 2(d) plot the BER performance of the uncoded transmission versus the disturber relative power, for BPSK, QPSK, 16-, and 64QAM respectively. For each constellation, the three curves correspond to SNR values of 10, 20 and 40 dB. At SNR equal to 40 dB, the channel thermal noise is negligible compared to the disturbing signal power. Most errors are thus due to the interfering signal. This explains the thresholds that are clearly visible; each constellation symbol originally transmitted is shifted by the amount of energy of the disturber constellation symbol scaled by the relative power of the disturber Pdist- As long

330

as Pdist is small enough not to shift the original symbol out of its maximum likelihood area, the errors observed are exclusively due to thermal noise. When Pdist increases, the original symbol is progressively shifted out to an adjacent symbol or even further, leading to an increasing number of erroneous bits. Consequently, there are up to 6 thresholds observable with 64QAM modulation, 4 with 16QAM, 2 with QPSK, and a single one for BPSK. Those thresholds correspond to an integer multiple of the Euclidian distance. For a SNR of 10 and 20 dB, the general shape of the curves are similar to the 40 dB case. However, the presence of higher power white noise smoothes

1

° r —. . .

r=

lOdBSNR 20dBSNR ••— 4 0 d B S N R

...

lOdBSNR 20 dB SNR 40 dB SNR

I;!:

:

1: i!:::=:

= :~ : E

4i

;:= '

•

•

-

1

.::::^|t::::::r: } ::;:.;! ::;;;::;;;::';;:i:f:i;i

ISlwg-:

ssr

,.H,

io- 4

(a) BPSK

(b) QPSK

z^=ir: :::=:3::::::::::

L'^i'*'*.-^*-"*'

::::::; :: = 3 ! * ^

:3-r}}i:=E

/| j (c) 16QAM

::::.;:::

.^,

10dBSNR • — 20dBSNR - - - 40 dB SNR

/

""" /:

lOdBSNR . . . . 20dBSNR . . . 40 dB SNR

(d) 64QAM

Figure 2. B E R of IEEE802.11a link versus t h e interferer relative power for the four constellations, with similar subcarrier modulation scheme for the disturbed and disturbing signals

331

Figure 3. BER of IEEE802.11a link versus the interferer power, for the four constellations, as a function of the disturber constellation (at 20 dB SNR)

out the effect of the steps caused by the disturber symbols. For the case of 10 dB SNR, we can clearly see the BER reaching a threshold: this characterize the situation where the link performance is not interference-limited but channel noise-limited. Finally, it can be noted that higher capacity modulation schemes are much more sensitive to inter-cell interference. For instance at a SNR of 20 dB, using 16QAM, a disturber of —4 dBc limits the minimum BER achievable to 3 x 10~ 2 , while —6.7 dBc is required to reach an identical value for the 64QAM scheme. 3.2

Synchronous disturber with a different constellation

In this section, we are interested in analyzing the disturbance of link performance as function of the disturber constellation. As in the previous set, it is assumed that both signals are totally synchronous. Figure 3 shows the BER performance of the link versus the relative power of the disturber for each of the four constellations, with a disturber using respectively BPSK, QPSK, 16QAM, and 64QAM modulation and at 20 dB SNR. Overall, it appears that the worst disturbance is endowed when the disturber is using QPSK modulation. On the other hand, for the three lower size constellations, the BPSKbased disturber provides minimal degradation of the link. For the 64QAM link, the disturber gives slightly worse results that the BPSK one. In all cases, the variation of performance degradation depending of the modulation used is rather narrow, only about 1 dB.

332

3.3

Asynchronous

disturber

The two previous analysis where assuming the case where both transmitter's OFDM symbols are synchronous in the time domain. This section reports the effect of a time synchronization offset between the disturber and the signal of interest. For simplicity, we assume here this difference is a multiple of the sampling period. Figure 4 plots the BER versus synchronization offset for all subcarrier constellations. The SNR is assumed to be 20 dB and Pdist is equal to —3 dB. The synchronization offsets range from 0 to 80 samples, which corresponds to the whole range of possible values, since in IEEE802.11a the total OFDM symbol length is 81 samples long. This explains the symmetry of the curves around At = 40. The curves obtained are clearly separated in two categories each grouping around the two low and high modulation schemes. First, for the 16QAM and 64QAM cases, the difference in BER versus At is very small and can reasonably be considered to be negligible. The worst degradation corresponds to the synchronous situation. Thus, the results obtained in the two previous sections can be used as worst-case scenarios. For BPSK and QPSK subcarrier modulation however, the situation is surprisingly totally opposite. While an extremely low BER is achievable with a synchronous disturber power of —3 dBc, a level of about 2 x 1 0 - 5 and 3 x 10~ 3 is obtained in the worse case where A t is equal to half of the OFDM symbol duration. Consequently, the curves obtained in the two previous cases for the two lower modulation schemes have to be considered as the best cases in regard to synchronization offset. BER ot IEEE BOB. 11a U n k u h m ^ o n ol tha ayndironaatkn ahrt

Figure 4. BER degradation versus synchronization offset, for all subcarrier constellations (at 3 dB SNR)

333

• -1.5 dBc interferer • - 1 dBc interferer • -0.5 dBc interferer • 0 dBc interferer

- T - - 3 dBc interferer •2 dBc interferer - * - - 1 dBc interferer — 0 dBc interferer

1C:requency offset A

Frequency offset A (a) BPSK link with BPSK disturber

(b) QPSK link with QPSK disturber

Figure 5. BER versus carrier frequency offsets for different disturber relative power (at 20 dB SNR)

Further simulations not reported here show a similar behavior for other values of the disturber power. 3.4

Disturber with carrier frequency offset

In this last section, we analyze the variation of the disturbance as a function of the frequency carrier offset between the two transmitters. Referring to the standard, the carrier frequency tolerance is of 20 ppm, which in the 5 GHz band is on the order of 100 kHz. This sets the maximum value of the carrier frequency offset Aj?. It is worth noticing that this is a rather large value, since the subcarrier spacing is equal here to 391 kHz. From an analytical point of view, the demodulated symbol is the sum of the original symbol, the channel noise, and the disturbing symbol shifted by its carrier frequency offset. The disturbing signal contribution can be further decomposed into the original transmitted symbol and a frequency offset generated noise-like component. Pollet et all has shown 6 that the power of this last component is shown to be proportional to the square of A p. Thus the disturbed link receiver sees an overall interference component scaled by Pdisp and divided into a disturbing symbol and a frequency offset generated random component proportional to the square of A jr. Consequently, a higher value of the frequency carrier offset

334

Af leads to a lower SNR and thus higher BER. Figures 5(a) and 5(b) plot the resulting BER versus the carrier frequency offset Ap with disturbing signal relative power as a parameter for BPSK and QPSK respectively. For both subcarrier modulations, it can be seen that for a given disturber power, the link BER remains at the same level whenever Ap is within the 250 Hz to 5 kHz range. As expected, the BER raises significantly for a higher frequency offset i.e. from 5 kHz to 100 kHz. This variation is especially important for lower values of Pdist a n d more significant for BPSK than for QPSK. Finally, similar simulations not reproduced here show that the carrier frequency offset has only very little influence on the link performance for 16QAM and 64QAM modulations. To reach for instance a BER of 10~ 2 with a link using a 16QAM subcarrier constellation, a variation of Ap from 0 to 100 kHz can be compensated by a variation of Pdist of about 0.4 dB. The carrier frequency offset influence for the 64QAM link is even too small to be evaluated accurately. 4

Conclusions and further research topics

We have reviewed the effect of a disturbing signal on the BER performance of an OFDM-based link. The effects of each of the parameters characterizing the disturbing signal have been studied in depth. This has permit in determining their relative importance and therefore prepare the ground for starting OFDM-based inter-cell interference performance simulations using a more realistic channel propagation model. From a system point of view, the total lack of control in the use of the unlicensed ISM band for high speed WLAN might lead to situation where serious interference can prevent the system to work properly, or at least to provide its maximum capacity. The problem is directly inherent in OFDM modulation and no simple technical solution can be reliably foreseen. This is especially true for higher modulation schemes, making them more risky to use in a dense cellular environment. In the cases where 64QAM and 16QAM schemes can not be used, the maximum capacity of the link is limited to 18 Mbits/s with 3/4 coding rate and 12 Mbits/s with the more robust 1/2 coding rate. This is quite comparable to the 11 Mbits/s that can be reached by the b version of the IEEE802.il standard 3 , this last one making use of a far more interference-robust spread spectrum modulation scheme. Consequently, it appears likely that the higher spectrum efficiency of OFDM modulation has been preferred to the robustness of spread spectrum techniques. However, this engineering decision might

335

have to be reconsidered more carefully if we reach the situation where a lack of bandwidth availability in the ISM band forces the high capacity WLAN modems to fall back to the rate equivalent of those in the previous generation. Finally, it is worth underlining that the results reported are conservative due to the assumption of perfect synchronization. A study of the performance degradation consequent to imperfect synchronization due to the presence of a disturbing signal might lead to much worst results. This might be investigated in future research work. Acknowledgments This work has been supported in part by the Academy of Finland (Grant for project 50624). References 1. IEEE Standardization body, "IEEE802.11-1999, Local and Metropolitan Area Network - Specific Requirements - Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications", 1999. 2. , "IEEE802.11-1999a, Supplement to IEEE802.11-1999, Part 11: Wireless medium access control and physical layer specifications: high speed physical layer in the 5 GHz band", 1999. 3. , "IEEE802.11-1999b, Supplement to 802.11-1999,Wireless LAN MAC and PHY specifications: Higher speed Physical Layer (PHY) extension in the 2.4 GHz band", 1999. 4. ETSI commitee, "Digital Video Broadcasting; Framing structure, channel coding and modulation for digital terretrial television", European Telecommunication Standards Institute, August 1997. 5. , "ETSI TR 101 683 - Broadband Radio Access Networks (BRAN); Hiperlan Type 2; System overview", European Telecommunication Standards Institute, February 2000. 6. T. Pollet, M. Van Bladel, and M. Moeneclaey, "BER sensitivity of OFDM systems to carrier frequency offset and Wiener phase noise", I E E E Transactions on Communications, Feb.-Mar.-April 1995, vol. 43, no. 2, pp. 191-193.

Part 12 Mobile Ad-Hoc Networks

A N A L Y S I N G C A P A C I T Y I M P R O V E M E N T S I N WIRELESS N E T W O R K S B Y RELAYING H. KARL, S. MENGESHA Telecommunication Networks Group Technische Universitat Berlin Einsteinufer 25, 10587 Berlin, Germany E-mail: {karl,mengesha}8ee. tu-berlin. de In infrastructure-based wireless networks, the capacity achievable at an access point is an important metric for system performance. In this paper, we present an analytical investigation of the use of relaying in wireless cells in order to improve the total goodput at an access point. A simple scenario of two cells is considered, formulas for SINR, packet error rate, and goodput are derived and numerically evaluated for both the standard direct communication and the communication via relaying terminals. We will show that the path loss coefficient and the physical location of terminals are the most important parameters and that for realistic scenarios, relaying can improve the total goodput of a cell by about 60%.

1

Introduction

In wireless networks, the available bandwidth for transporting traffic is a scarce and valuable resource. In outdoor environments, this has been exemplified by the large costs of UMTS frequency bands. Given a fixed amount of spectrum, maximizing the capacity for successfully transporting user traffic is, hence, a question of great practical importance for both outdoor and indoor environments. Many different approaches have been used to improve the capacity (e.g., coding and modulation or appropriate medium access techniques) and recently, the use of relaying has been considered not only to extend the coverage of a wireless network, but also for capacity improvements. One scenario is a network of access-point-based cells, where the question of improving the achievable goodput at such access points is an interesting metric. The fundamental impediment of large capacity is errors in the transmission. One of the main causes of transmission errors is interference, and reducing interference represents one reason why relaying could improve capacity: In a direct communication scenario, every terminal is communicating directly with the access point. While this does not cause problems for terminals that are close to the base station, terminals that are far away (called "far" terminals) need to use an overproportionally large amount of transmission power (due to the non-linear decay of received power over distance). This large transmission power, in turn, will create interference in the neighboring cells

339

340

(using the same or adjacent frequency bands). Even though such cells will be some distance away (using typical frequency reuse schemes), the fact remains that reducing the transmission power of terminals close to the edge of a cell holds a big promise for reducing interference and, hence, improving capacity. However, reducing transmission power results in a higher error rate when attempting to communicate directly with the access point. Communication with a terminal positioned in between should still be possible. Such an intermediate terminal could serve as a relaying point. But this relaying terminal now has to transmit a higher traffic load, which is in general not possible at the curent modulation: A "faster" modulation with a higher bit rate needs to be used. Normally, a higher bit rate would incur the penalty of higher error rates at the same level of noise and interference, but as the level of interference is reduced (far terminals reduce their power), this might actually be feasible. It is this tradeoff between lower interference by reduced transmission power at the border of cells on the one hand and increased error rate due to the need to increase data rate in the interior of a cell on the other hand that this paper investigates. The crucial point is selecting transmission power and, in particular, modulation schemes for both far and relaying terminals. Appropriately choosing modulation types depending on interference level and relaying task is the novel approach suggested in this paper. For a simple model, the total goodput of a cell is analyzed and numerically evaluated. As the main contribution, we will show that for many relevant scenarios it is possible to choose transmission power and modulation for a relaying approach such that capacity can be improved up to 60% over direct communication. After considering some other work related to capacity increase in wireless networks in Section 2, the model itself is introduced in Section 3. As the present work is part of a project 1 that uses HiperLAN/2 as a case study of a wireless network technology, this analysis here uses a simple time division multiple access model. The goodput achievable in this model using both direct and relay communication is analyzed in Section 4 and numerical evaluations are shown in Section 5. The results are discussed and conclusions are drawn in Section 6, which also outlines prospects for future research. 2

Related Work

Relaying in ad-hoc networks has received considerable attention in the research literature; only a few important papers can be highlighted here. A more detailed discussion can be found in an extended version of this paper. 2 SHEPARD showed a decentralized channel access scheme for a scalable packet radio network that is free of packet loss. 3 Routing is used between

341

terminals to relay data, and interference reduction is an optimization goal for the routing decisions. These basic ideas are present in our work as well, however, the main difference is that in our scenario, relaying terminals are assumed to have traffic of their own to transmit and that the modulation and hence the intereference susceptibility can be varied. With regard to the capacity of wireless networks, GUPTA and KUMAR 4 have studied the capacity of randomly located ad-hoc networks. Their result shows that for n identical, randomly located nodes, the maximum achievable throughput under optimal circumstances is - ^ times the transmission rate and as the number of nodes n per unit area increases, the throughput decreases accordingly. GROSSGLAUSER and TSE 5 used a similar situation as in 4 but introduced mobility into the model and showed how the capacity of this wireless ad-hoc network can be increased using a single mobile relay node. However, the increase of the capacity is at the cost of considerable delay. Recently, BRONZEL et al. 6 have also shown how the capacity of a singlerelay network (called "single hop relay" in their paper) can be increased by reducing the interference in a CDMA network. While their model is somewhat similar to ours, the main difference is in the TDMA vs. CDMA choice, as well as in some conclusions drawn about the impact of the path loss coefficient. 3

Model Definition

For an analytical investigation of relaying, a very simple model is used: two cells, each with an access point and two terminals. Our goal is to check if and under which circumstances the total goodput (the amount of successfully transfered bits from both terminals to the access point per unit time) is increased when relaying is used. A symmetric layout of access points and cells has been chosen (Figure 1): six nodes are located on a single line, d is the distance between terminals within a cell, the distance between the two outmost terminals of each cell is D (the separation of the cells). All terminals use the same communication channel, as we are interested in improvements due to reduced interference. The analysis will only consider the uplink communication from mobile terminals to access points. The inner terminal (Mi or M4) use transmission power PT (in Watts), the outer terminals (M2 and M3) use transmission power PT (how to actually choose PT and PT is discussed in Section 5. The path loss model relating transmitted and received power PT and P R is simple and depends only on distance x and path loss coefficient a: P R = ^ . The actual packet error rate results from the received power P R , the noise

342

rr\ r>r. r^-. W (M?» (£) (A) '^C»71 Y •-<•A vY X... *—fYX

d

d

/

D \

AK.

..L...

-"~r-A Y" d

d

Figure 1. Scenario description

N (in Watt) and the interference I from other transmissions in the same channel, constituting the signal to noise and inference ratio SINR = PR/(N +1). In this model, adjacent channel interference is ignored. Interpreting the SINR as carrier-to-inference ratio C/I, the mapping to packet error rate (PER) described in 7 for HiperLAN/2 can be used. A simple exponential curve fitting is used to mathematically describe the mapping. The PER is then 8 : PER(s,i)

= 10ai°*+biS+Ci

(1)

where s = 10 log 10 SINR (SINR in dB), i is the modulation index (i = 1 , . . . , 7 corresponding to nominal bit rates (NBR) between 6 and 54 MBit/s) and a*, bi, and q are modulation-dependent parameters. 8 The PER is normalized to stay between 0 and 1. While a more general error model is desirable, we believe that this approximation is close enough for some initial insights. As we are interested in the use of time-division multiple access techniques, a communication schedule needs to be defined for the direct a well as the relaying case in both cells. Section 4 will described the schedules in detail. As a worst-case scenario, we are here considering the simultaneous occurrence of the uplink phase in both cells. It should also be mentioned that we are only taking into account the actual data transmission phase; overhead needed to construct schedules (and to determine relaying stations) is not considered. Finally, all terminals are assumed to transmit at the maximum possible data rate to the access point, by only considering the nominal capacity of a given transmission. Otherwise, the net bandwidth would correspondingly be reduced due to protocol overhead. Intra-cell traffic is not also considered. This traffic assumption is actually conservative as it puts considerable requirements on the relaying node, which has not only to transmit its own traffic but also the other node's traffic. This results in the need to use faster modulations between inner terminal and access point (see Section 4) with inferior error behavior. Under lower load, this could be avoided, which should in turn improve the performance of relaying.

343 Table 1. Communication schedules

Slot 1 Slot 2

Cell 1 Mi -> A\ Mi - • A\

Cell 2 Mi -> A-i. M\ -> A
(a) Direct communication

4 4-1

Slot 1 Slot 2

Cell 1 M\ -)• A\ (fast) M2 -»• Mi

Cell 2 M 3 - > M4

M 4 -> A 2 (fast)

(b) Relay communication.

Analysis Direct communication

Initially, a communication schedule needs to be determined in order to facilitate the computation of the interference in the two cells and the goodputs at the access points. The uplink phase considered here is scheduled using two time slots, one for the inner and another for the outer terminals. To keep the symmetry between the two cells, inner and outer terminals of the two cells operate in alternate slots. These terminals, independent of each other, can use any possible modulation to transmit their traffic load to the access point. The resulting schedule is shown in Table 2(a). In the following derivation, only one cell is considered; result for the other one are symmetrical. Based on the schedules, the actual SINR values can directly be computed as S I N R M I - ^ ! = N+p^/(2d+D)a f° r transmitting from Mi to A± and as SINRM2-y>i1 = jv+p^/(3d+D)^ for transmitting from M 2 to Ai. Once the SINR values of each transmission is computed, the PERs can easily be determined from Equation (1) for the the employed modulation type i which is considered to be identical for both terminals. The goodput GP for a particular terminal can, then, be approximated as GP = (1 — PER)NBR. However, since the two terminals have to share the channel with each other, the nominal data rate needs to be divided by two (without considering a downlink phase, which would further scale down the goodput). Thus, the goodput is (using the PER Equation (1)): = (1 - PER(101og 1 0 (SINR M l ^^ 1 ),i))(NBR i /2)

(2)

G P C O K ^ = (1 - PER(101og 1 0 (SINR M 2 ^ 1 ), j))(NBR,/2)

(3)

GP(i)&

Cell l's total goodput is the sum of goodput of terminal 1 and 2.

344

4-2

Relay

communication

The schedule for this setup is slightly different: on the one hand, only two time slots are used to communicate the traffic of both mobile terminals to the access point within the cells. On the other hand, the relaying stations Mi and M4 should transmit both their own traffic as well as the relayed traffic of M 2 and M3, respectively. Evidently, it is not possible to use the same duration of time and the same modulation as in the previous case (when both stations transmit at a bandwidth close to the maximum possible bandwidth of a chosen modulation). Hence, to transmit from the relay station to the access point, a "faster" modulation has to be used. Here, we assume that for this communication link, a modulation which is at least faster than that of the communication between the outer mobile terminal and relay terminal is used. Thus, the actual schedule looks like in Table 2(b). This schedule is actually a pessimistic approximation. The relay terminal is always using the faster, more error-prone modulation, whether or not the outer terminal has sent any data that is to be relayed to the access point. In the particular traffic model assumed here, the outer terminal always has data to sent which is unlikely to be the case in realistic scenarios. Hence, adding a mechanism that uses faster modulation between relay and access point only when necessary should again improve the error behavior of this link. Such a mechanism, however, could imply additional signaling overhead and delaywe leave this problem for further study. Based on the schedule, the actual SINR values for transmitting from Mi to Ai and from M2 to Mi can be computed directly as S I N R M J - ^ I = A f + p f / f i W aiid S I N R M ^ M ! = N+P^/IU+D). , respectively. The main difference between the direct and the relay case is now apparent: lower transmission power is used close to the border of a cell, improving the interference situation, yet on the downside, faster and more error-prone modulations need to be used close to the interior of a cell. Determining PER and goodput is also different from the direct case as two different modulations going over the different hops are used. Consider first the goodput for terminal Mi, which uses modulation i with nominal bit rate NBRj to transmit to the access point. Since Mi is transmitting only for half the time, NBRj has to be divided by two. Moreover, it is only part of the data that belongs to Mi in the given time slot. Thus, the actual goodput of Mi for transmitting its own data is G P ^ W M ) = (l-PER(101og 1 0 (SINR M l ->^ 1 ),i))(NBR i - NBR,)/2 (4) where j is the modulation index used by Mi to relay its data to Mi.

The

345

goodput of M 2 depends on both the NBRj it is using when communicating with Mi and the NBR* that Mi is using to communicate with the access point. Similar to the direct case, M 2 can transmit only half the time and hence its NBR is reduced by half. The bandwidth from Mi to Ai should not limit the goodput of M 2 , since we already assumed that NBR* > NBRj. More complicated is the error handling: a packet from M 2 needs to be transmitted successfully between M 2 and Mi as well as between Mi and Ai. Hence, the goodput for Mi is: G P S £ W * » i ) = (1 - P E R C l O l o g ^ C S I N R M ^ ^ i ) ) * (1 - P E R U O l o g ^ S I N R M ^ M j , j ) ) ^

(5)

Just like in the direct communication case, these numbers would have to be downscaled correspondingly for inclusion of a downlink phase. However, as this needs to be done in both case, this scaling factor cancels out when looking at the ratio of the total goodput in both cases and is hence not considered here. The total goodput at the access point is then the sum of the individual terminal goodputs in the corresponding cell.The optimization problem is hence, for both direct and relay communication, to choose modulation types and transmission power such that GPi + GP 2 is maximized, for any given scenario (physical layout, path loss, etc.). 5

Numerical Evaluation

We have examined the capacity of both direct and relay communication using numerical solutions of the equations derived in Section 4. The numerical solution obtained depends on the physical layout described by d and D, on the transmission power PT and PT, path loss coefficient a, the noise level N and modulation type. Our evaluation is, thus, based on the SINR, the PER, and the goodput at both near and far terminal, using either direct or relay communication. One representative example of such results is shown in Table 2. The figures in the table are the optimal goodputs for the given parameters a, N, d, D and transmission powers PT and PT. Figure 2 gives an overview of the ratio of the optimal total goodput in a cell between direct and relaying communication for various values of a. Ratios with values larger than 1 indicate a better relaying performance. For both communication paradigms, all terminals use the best combination of modulation types and transmission powers in order to achieve a maximum total goodput. As can be seen from the graphs, the capacity ratio improves

346 Table 2. Goodput in MBit/s for different scenarios.

SINR PER Goodput Total goodput

Mi - • A\ M-i. -> A\ 16.1633 4.7671 0.0026879 0.461296 3.23222 5.98387 9.21609

(a) Direct communication, Pj- = 0.1W, P'T = 0.05W, Q = 4, N = 1.64 x 1 0 - 1 3 W , d = 75m, D = 10m and 12MBit/s modulation.

SINR PER Goodput Total goodput

M\ ->• A\ Mi -> A\ 16.1633 16.8083 0.2822 0.00024975 10.767 2.15286 12.9199

(b) Relaying communication, P y = 0.1W, P'T = 0.05W, a = 4, N = 1.64 x 1 0 _ 1 3 W , d = 75m , D = 10m, M 2 -> Afi: 6MBit/s, Mj -> A i : 36 MBit/s.

with increasing a. For ct = 4, for example, an improvement of about 60% is achieved when the terminals are far away from the access point. Evidently, there are placements of terminals where direct communication is superior and others where relaying communication is better. It is also possible to conclude that relaying is particularly beneficial when considering scenarios where mobile terminals are far away from their access points, and also for the case of cells located close to each other. While the last case is usually avoided with spatial reuse, it might occur more often when relaying within a cell is considered using different channels (requiring a larger number of frequencies than might be available in usual spatial reuse patterns). 6

Conclusions and Future Work

The analysis and numerical evaluations in Section 5 have shown that relaying can achieve considerably larger total goodput than direct communication in a TDMA-based wireless network. Critical parameters are the path loss coefficient a and the physical location of terminals. In our scenario, best results were achieved when terminals were far away from their access points, yet the far terminals were still comparably close to each other; access points were still separated by a considerable distance, as would be required by a spatial reuse scheme.Best results shown here were achieved for a = 4. In fact, values for a between 3 and 4 are commonly found in indoor environments, making relaying attractive for realistic scenarios. It is also possible to look at this result under a different perspective: The particular strength of relaying is a setup where cells are comarably large, yet close together. Hence, if a certain goodput is required per cell, it is

347

D[m]

" °

D m

d[m]

I )

(a) a = 2.5

" "

d[m]

(b) a = 3

D m

I l

d[m]

(c) a = 4

Figure 2. Ratio of total goodput for various alpha

possible to place such cells closer together when relaying is used, as the spatial "protection" between cells need not be as large as with direct communication. This should prove particularly beneficial in scenarios where a lot of cells should be placed within a limited amount of space, e.g., office builldings, exhibitions halls, etc. As this analysis only attempted to probe the potential of relaying in a very simple scenario, there are many ways to extend this work. Currently, we are considering the downlink case in a relaying scenario, and preliminary results are encouraging. Controling the relative share of goodput per terminal

348

in a fair manner, extending the work to larger scenarios (more terminals per cell and more cells), and adding more terminals (instead of improving goodput per terminal) are our intended next steps. However, as this will become more and more complex to analytical analysis, we are currently developing a simulation environment to investigate both capacity and energy efficiency issues for relaying communication. This simulation will also take into account the impact of ARQ protocols, scheduling of communication to reduce delay when relaying, power adaptation schemes, and signaling overhead for protocols to determine terminals suitable for relaying. Also, the use of multiple channels within a single cell (along with proper physical channel models) is intended. A long-term goal of our research is to devise mechanisms to automatically detect situations when it is beneficial to turn on relaying. References 1. Webpage of the IBMS 2 project, http://www.ibms-2.de. 2. H. Karl and S. Mengesha. Analysing capacity improvements in wireless networks by relaying. TKN Technical Report TKN-0103, May 2001. http://www-tkn.ee.tu-berlin.de/publications/ tknrreports.html. 3. T. J. Shepard. A channel access scheme for large dense packet radio networks. In Proc. of Applications, Technologies, Architecture, and Protocols for Computer Communications (ACMSIGCOMM), pages 219-230, Palo Alto, CA, August 1996. 4. P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Trans, on Information Theory, November 1998. 5. M. Grossglauser and D. Tse. Mobility increases the capacity of wireless adhoc networks. In Proc. Infocom, 2001. 6. M. Bronzel, W. Rave, P. Herhold, and G. Fettweis. Interference reduction in single hop relay networks. In Proc. of 11th Virginia Tech Symposium on Wireless Personal Communications (to appear), Blacksburg, VA, June 2001. 7. J. Khun-Jush, G. MalmgrenyP. Schramm, and J. Torsner. HIPERLAN type 2 for broadband wireless communication. Ericsson Review, 2:108119, 2000. 8. T. Osche. HiperLAN/2 PHY layer simulation in C + + . Studienarbeit, Technische Universitat Berlin, Fachgebiet Telekommunikationsnetze, Berlin, Germany, July 2001.

WIRELESS LAN WITH WIRELESS MULTIHOP BACKBONE NETWORK (WMLAN) KENICHI MASE, NAOYUKIKARASAWA, MIHO KUSUMI, KEISUKE NAKANO, AND MASAKAZU SENGOKU Department of Information Engineering, Niigata University, Japan E-mail: [email protected] This paper presents a novel concept and system for wireless LAN, that is, Wireless multihop LAN (WMLAN). This network is just the wireless LAN for mobile stations. Its backbone is the wireless multihop network, where mobile ad hoc network (MANET) technologies are used to support communication between access points. The benefits and applications of this network system are discussed. Two architectures realizing WMLAN are presented. An experimental system is developed to evaluate network performance.

1

INTRODUCTION

Number of mobile communication subscribers continues to grow rapidly in many counties and has even exceeded that of conventional telephone subscribers in some countries. These statistics show how mobile communication fits to the needs in the present days. It is no doubt that mobile communication services will be enhanced in 21st centuries, which requires further innovations in mobile communication technologies. Two typical mobile communication systems, currently used, are cellular mobile and wireless LAN (WLAN). The former covers nation-wide as well as global communication, while the latter covers local communication. WLAN is used only indoors in many cases. However, it can also be used outdoors. For example, WLAN can be used for local communities. Cellular mobile communication services are usually planned and provided by mobile communication service providers. On the other hand, WLAN services can be planned and provided by private sectors without help of mobile communication service providers. For example, city office may plan and install WLAN for the local community. Even volunteers in the community may do so. In this case, communication in the local community can be free of charge, unlike cellular mobile communication services. The system without charging can economically be operated. This kind of mobile communication service may be attractive in the local community. The backbone of the WLAN may be constructed of either wired or wireless networks. We focus on wireless backbone, considering wiring in a large area outdoor is generally expensive. In particular, we propose WLAN with wireless multihop backbone network. In the following, we call this network concept as

349

350

Wireless Multihop LAN (WMLAN). In WMLAN, a message is carried from a source access points (APs) to a destination AP by repeating wireless hopping at intermediate APs on the path. Wireless multihop is used in the mobile ad hoc network (MANET). In IETF (Internet Engineering Task Force), there is a group called MANET (Mobile Ad Hoc Networks) [1], which mainly studies routing protocols for ad hoc networks. Ad hoc networking is a technology to instantly make a network for a specific and temporary purpose within a group. Typical ad hoc network applications include military affairs, events and disaster recovery. It should be noted, however, wireless multihop can be used for wider applications beyond mobile ad hoc networks in realizing a novel communication environment, which may not properly be covered by conventional cellular mobile technologies [2]. Wireless backbone for WMLAN is one of the examples. We assume that APs are placed to cover the service area. Each AP has at least one neighbor AP, which is within transmission range. In this case, we say there is a link between the APs. We assume that each AP has at least one path to any AP, where adjacent APs on the path have the link between them. AP placement and connectivity assurance are important issues in designing conventional WLAN systems. We expect that these issues may become simpler by using MANET routing protocol in WMLAN. WMLAN concept includes integration of WLAN and MANET technologies. We can consider two types of integration, horizontal and vertical. In the former, WLAN and MANET both have their users (mobile stations), and a user in WLAN can communicate with a user in MANET. Our proposal can be regarded as the vertical integration, where the user interface is provided by WLAN and MANET is transparent from users. WMLAN may be useful to realize a novel information infrastructure for local communities with fixed APs as well as ad hoc communication environment on occasions such as big festivals and events with portable APs. The rest of the paper is organized as follows: Sec. II presents the concept of WMLAN. Sec. Ill presents two WLAN architectures. Sec. IV describes experimental systems and results of experiments. Sec. V gives concluding remarks. 2

THE CONCEPT OF WMLAN

An image of WMLAN is shown in Fig.l. APs provide WLAN service for the mobile stations in the coverage area. At the same time, they are connected to each other using wireless multihop and work as MANET. Accordingly, WMLAN architecture includes two layers, WLAN layer and MANET layer (Fig.2). In WLAN layer, mobile station-AP interface is same as that of WLAN. In MANET layer, APs are regarded as the mobile host, even if APs may be stationary. Since we use MANET routing protocol, there is enough freedom in placing APs. The minimum

351

Figure 1. An image of WMLAN. MS

MS WLAN Layer

MANET Layer

<

•

*H h *! t ^ - _fcns;_r .„

._

M S : Mnhile Station

AP: Access Point

Figure 2. An high-level view of WMLAN architecture.

requirement in AP placement is that connectivity is maintained between any AP pairs through multihop path (no partition). Routing protocols for MANET are classified into two classes, table-driven routing and on-demand routing. In WMLAN, AP mobility is usually zero or very low. Thus, on-demand routing protocol seems not necessary. In table driven routing, routing information is periodically distributed in the MANET. Relatively long interval can be allowed in WMLAN due to low AP mobility. Thus, routing overhead in WMLAN may be negligible. Comparison of typical mobile network systems is shown in Table 1. WMLAN has remarkable characteristics compared with conventional mobile network systems such as WLAN, MANET, and cellular networks. WMLAN is superior to WLAN in terms of coverage and deployment speed due to freedom in AP placement and connections, which is brought by ad hoc routing technologies. It is superior to MANET in terms of connectivity and coverage due to the use of APs. It is superior to cellular mobile networks in terms of deployment speed. On the other hand, WMLAN has limitation in terms of coverage compared with cellular mobile networks.

352 Table 1. Comparison of mobile network systems.

WMLAN AP/BS Connectivity Coverage Deployment Speed Deployment Cost

Required High Medium Quick Very Low

WLAN Required High Low Medium Low

MANET Cellular Networks Not Required Required Low High Very Low High Instant Slow Negligible High

AP: Access Point, BS: Base Station When we turn our attention to the mobile station, there is no difference between WMLAN and WLAN. The mobile station for WLAN can also be used in WMLAN. This is an advantage of WMLAN for WLAN users, compared with MANET and cellular networks. These systems are not always alternatives, and each of them may have suitable areas and circumstances to be applied. Some of them are complementary and can be combined to provide total and beneficial communication services. By applying WMLAN technology to a local community, a community network can easily be created. APs may be placed along the avenues and streets so that the adjacent APs are within the transmission range to each other. The size and coverage of the network can easily be adapted to meet varying community demands, by simply adding or deleting APs. Note that a conventional community network is targeted to realize a closed network for a specific area. Such a network is not open for visitors to use. The coverage of the network is fixed and not extendable beyond the area due to physical constraints. WMLAN technology is adequate to realize an open community network without such barriers. In an open community network, loudspeakers may become useless for guides in public places. For example, sales information is delivered to pedestrians in the shopping center, or to the residential area using the open community network, without making a noise. Number of traffic accidents significantly decreases, if necessary alarm information is spread from sensors embedded in traffic signals at intersections, or in light poles along streets to pedestrians and drivers in the neighborhood. This system is also useful to navigate old people and the disabled, walking along the street. Other benefits include crime prevention, electric power saving, living environment improvement. In addition, an open community network is useful to keep communication between victims and rescue workers as well as among rescue workers in emergencies and disasters. These are only a few examples and an open community network may have various applications in the future [2]. The WMLAN-based open community network may have one or several gateways to the Internet. Community members do not need to have Internet access by themselves. WMLAN can cover wider areas and aggregate more Internet access traffic than WLAN does. Thus, WMLAN is useful to efficiently construct the socalled hot spot networks.

353 3 3.1

WMLAN ARCHITECTURES LAN Emulation (MA C over IP)

The protocol stack of WMLAN based on LAN emulation is given in Fig. 3. Extended service set (ESS) is supported in WLAN layer. Assume a mobile station of a basic service set (source BSS) sends a frame to another station, which belongs to a different BSS (destination BSS). The source station sends a frame to the AP (source AP), to which it is associated. The source AP receives the frame, which includes the MAC address of the destination station. The distribution system (DS) in the WLAN layer identifies the MAC address of the AP (destination AP), to which the destination station is associated. The source AP encapsulates the received frame in an IP datagram and sends this datagram to the destination AP by using MANET routing protocol. The destination AP decapsulates the IP datagram, receives the frame, and forwards it to the destination station. The source AP needs to resolve IP address of the destination AP based on MAC address of the destination AP. For this purpose, each AP maintains destination address resolution table, which is a list of destination AP MAC address and destination AP IP address. The IP address of the destination AP is thus obtained from the table based on the destination AP MAC address. The destination address resolution table must be updated coping with addition and deletion of APs. MANET routing protocol may be modified to do so. The source AP and destination AP interface is same as that in WLAN. Thus, MANET provides tunnels for WLAN frames. In other words, MANET layer provides LAN emulation (MAC over IP) service to WLAN layer. MANET is transparent from mobile stations. MANET IP datagram is encapsulated by MAC frame, when it is sent from AP to AP through wireless links. In order to avoid unnecessary fragmentation, the frame length in WLAN layer should be limited to (MTU - x - y), where x is IP header length and y is frame header length. MS

AP

AP

AP

MS Application

Application

TCP/IP

TCP/IP LLC

Bridge Relay

Bridge Relay

LANE IEEE802.il (MAC)

IEEE802.il (PHY)

IEEE 802.11 (MAC)

IP

IP

LLC

LLC

1KEE802.1 ] (MAC)

lfeEES02.il (MAC)

IEEE802.U EEE802.U (PHY) (PHY)

LLC

LANE

IEEE802.il (PHY)

IEEE 802.11 (MAQ

IEEE802.il ffiEE802.ll (PHY) (PHY)

L_Z Figure 3. A protocol stack of WMLAN based on LAN emulation.

IEEE802.il (MAC)

IEEE802.il (PHY)

354

3.2

IP-in-IP encapsulation

The protocol stack of WMLAN based on IP-in-IP encapsulation is given in Fig. 4. BSS is supported in WLAN layer. Assume that a mobile station of a BSS sends a frame to another station, which belongs to a different BSS (destination BSS). The source station selects the AP of the source BSS (source AP) as the default route. A packet arriving at the source AP is sent to the AP of the destination BSS (destination AP), by using the IP-in-IP encapsulation. The destination AP receives the packet, decapsulates the packet and forwards the inner packet to the destination station. To support station mobility during communication, IP address of a station should not be changed due to mobility. The source AP needs to resolve IP address of the destination AP based on IP address of the destination station. For this purpose, mobile IP can be used. MS

AP

AP

MS

AP

Application

Application

TCP/UDP

TCP/UDP IP LLC

LLC

IP

IEEE802.il (MAC)

IEEE 802.11 (MAC)

LLC 1EEE802.11 (MAC)

IEEE802.il (PHY)

1

IEEE802.il IEEE802.M 0>HY) (PHY)

1

IP

IP

IP

1

IP

LLC

LLC

LLC

LLC

1EEE802.I1 (MAQ

IEEE802.11 (MAC)

IEEE 802.11 (MAC)

IEEE802.il (MAQ

IP

EEE802.1I (PHY)

J

l_

IEEE802.il IEEE802.il (PHY) ffHY)

1

1

IEEE802.il (PHY)

1

Figure 4. A protocol stack of WMLAN based on IP-in-IP encapsulation.

3.3

Comparison

Mobility management is completely treated in WLAN layer in LAN emulation. That is, the ESS hides the mobility of the mobile stations from everything outside the ESS. This solution requires standardization of inter-access point protocol (IAPP). On the other hand, mobility management is treated in IP layer on top of WLAN layer in IP-in-IP encapsulation. Efficiency in dealing with station mobility may depend on WLAN and Mobile IP technologies, WMLAN coverage, and station mobility. Note that MANET routing and destination address resolution is completely separated in WMLAN architecture. Station mobility is completely processed in the DS in case of LAN emulation or in the mobile IP in case of IP-in-IP encapsulation, and is not concerned with MANET routing protocol. In WMLAN architecture, AP may be portable, but mobility is usually low, while station mobility may be high. Thus MANET routing may be designed based on low mobility assumption. On the other hand, station mobility management should be designed based on high mobility assumption.

355

4

4.1

Experiments and results

Experiments

We are under development of experimental systems using the IEEE 802.11b WLAN. In our experimental system, an AP for the IEEE 802.11b WLAN is connected to a PC with Linux 2.2.12 to emulate an AP for WMLAN. We have analyzed the implementation issues of both approaches, LAN emulation and IP-in-IP encapsulation. The problem in the former is that we cannot control the MAC header's content, when the PC receives and sends data. That is, when the PC receives data from another PC and relays the data to the mobile station, the sender address is set to the address of the wireless interface of the PC. Therefore complete LAN emulation is difficult. We thus decide to take the latter approach, IP-in-IP encapsulation. In this case, PC supports encapsulation functionality as well as MANET routing and broadcasting functionalities. Currently, we focus on the multihop backbone network of WMLAN. We implement Mobile Mesh protocol [3]-[5], which is a table-driven MANET routing protocol, in PCs emulating APs. We have done several experiments to investigate throughput of the multihop backbone network. Experiment 1: Five PCs with IEEE 802.11 WLAN card are placed in line (Fig.5). Only two adjacent PCs are within direct transmission range to each other. Netperf is used to measure TCP Reno throughput. Maximum segment size is set to 1460 bytes. Each PC can transmit data to the adjacent PC with 5 Mbps throughput, when receiver buffer size is sufficiently large. Receiver buffer size is selected as the parameter. Throughputs from PC A to PC B (1 hop), to PC C (2 hop), to PC D (3 hop), or to PC E (4 hop) are measured for 10 sec. Ten measurements are conducted for each scenario and collected data is averaged over those measurements.

Figure 5. Experimental system.

356

Experiment 2: We next evaluate application level QoS using gnutella. The gnutella is a P-to-P application for sharing files and would be one of the attractive applications in WMLAN environments. We measure download time of 1 MB file in the network model in Fig. 5. Again, receiver buffer size is selected as the parameter. The file at the PC A is downloaded from PC B (1 hop), from PC C (2 hop), and from PC D (3 hop), and the file download times are measured. Ten measurements are conducted for each scenario and collected data is averaged over those measurements. 4.2

Results

Experiment 1: The relation of TCP throughput and number of hops is shown for different receiver buffer sizes (bytes) in Fig. 6. TCP throughputs for buffer size 5840 and 32120 are almost same. Throughput significantly decreases, when number of hops is increased from one to two in these cases. This sharp reduction is due to the hidden and exposed terminal problems in CSMA/CA environment. TCP throughput for buffer size 1460 is quite different. The TCP throughput in this case is about half of other cases, when number of hops is one. However, reduction of throughput when number of hops increases from one to two is not so sharp, compared with other cases. Accordingly, throughput is about double of other cases, when number of hops is more than 1. Packets are sent and received at almost uniform interval when buffer size is 1460 as shown in Fig. 7. Packet transmission interval is about 15 ms, This time is about same as that required for a packet to proceed 4 hops. Therefore, the hidden and exposed terminal problems are avoided. Thus, buffer size should be set to the same size as the maximum segment size to achieve highest throughput in this multihop communication environment. The similar problem is identified in [6], based on simulation study. This paper suggests the maximum TCP window size as the parameter to improve throughput. We have shown that buffer size can also be used to improve throughput based on the experiment as mentioned above. Unfortunately, our method may not work well when number of hops between the source and destination is more than 4. In these cases, packet transmission interval may be more than 15 ms, resulting unnecessary low throughput. The similar problem is also expected in the method using the maximum TCP window size. Rate control may be effective to improve throughput in general cases.

357

2

3

Number of hops Figure 6. Throughput versus number of hops.

3

4

5

6

7

8

10

TCP packet number Figure 7. Packet arrival time when number of hops is four.

Experiment 2: The relation of file download time and number of hops is shown for different receiver buffer sizes (bytes) in Fig. 8. The file download time when buffer size is 1460 is about double in lhop, but about half in 2 and 3 hops that when buffer size is 32,120. These results are consistent to the results in experiment 1. Thus. Application-level QoS is improved by choosing buffer size in these situations. However, improvement may not be expected when two or more PCs simultaneously request file download. Again, rate control in TCP may be effective to deal with this problem.

358

1

2

3

Number of hops Figure 8. File download time using gnutella.

5

Conclusions

This paper presents a novel concept for wireless LAN, that is, Wireless Multihop LAN (WMLAN). This network is just the wireless LAN for mobile stations. Its backbone is the wireless multihop network, where the mobile ad hoc network (MANET) technologies are applied to support communication between access points. The benefits and applications of this network system are discussed. Two architectures for WMLAN are presented, LAN emulation and IP-in-IP encapsulation. An experimental system based on IP-in-IP encapsulation is under development to evaluate network performance. Preliminary results of the experiments are summarized. Further study areas include end-to-end QoS and performance evaluation and development of rate control mechanism in TCP for WMLAN. References 1. 2.

3. 4. 5. 6.

http://www.ietf.org/html.charters/manet-charter.html K. Mase, M. Sengoku, and S. Shinoda, "A perspective on next-generation ad hoc networks-A proposal for an open community network-," IEICE Trans. Fundamentals, vol. E84-A, no.l, pp.98-106, January 2001. http ://www.mitre. org/tech_transfer/mobilemesh/ K. Grace," Mobile Mesh Link Discovery Protocol", Internet-draft, draft-gracemanet-mmldp-OO.txt, 2000. K. Grace," Mobile Mesh Routing Protocol", Internet-draft, draft-grace-manetmmrp-00.txt, 2000. S. Xu and T. Saadawi, "Does the IEEE 802.11 MAC protocol wirk well in multihop wireless ad hoc networks?," IEEE Communications Magazine, Vol.39, No.6, 2001.

AUTHOR INDEX

AbeK. Agarossi L. Andrew L. Arbaugh W. Babanskaja I. Bechler M. Chen B. Chui T. Chung K. Deneire L. Dombrowski K. Elbassioni K. Engels M. Fawcett J. Frankenfeldt H. Furukawa H. George L. Gerla M. Giangaspero L. Hadzi-Velkov Z. Hosseini H. Howie D. Huang P. Hurler B. Jamhour E. Jamin A. Jang K. Jappinen P. Jiang C. JinX. Kallmann V. Kamel I. Kampouridou S. Kapoor R.

113 253 177 131 145 47 79 187 241 285 145 79 285 3 145 154 228 273 317 164 305 121 27 47 57 327 263 91 14 253 47 79 37 273

Karasawa N. KarlH. KimK. KimS. KimY. Koivisto A. Komaki S. Kraemer R. Kusumi M. Kwon T. Ishibashi H. Langendorfer LeeH. Lee J. Lee Y. J. LeeY. Lenders V. Li J. Mahonen P. Mase K. Matsuura T. Matthaei I. Mengesha S. Mercier A. Mercier G. Methfessel M Morishita H. Miner P. Mori T. Muheim M. Nakano K. Okada M. Okamura S. Okanoue K.

360

Paltenghi G. Passas N. Porras J. Ranasinghe R. Rohani B. Sauvola J. Scanlon W. Sengoku M. Shankar N. ShiY. Shoemake M. Skyrianoglou D. Sowerbutts W. Spasenovski B.

317 37 91 177 305 121 187 349 131 14 100 69 3 164

Sun J. Tang F. Thoen S. Wan J. Watanabe Y. WolfL. XieW. XuG. Yamai N. Yamazaki S. YehC. Zanella A. Zhen B.

121 285 285 131 154 47 14 14 113 154 197,215 273 263

World Scientific www. worldscientific.com 4866 he

ISBN 981-02-4826-1