INFOCOM 2003Return to Main MenuUnderstanding TCP fairness over Wireless LANSaar Pilosof1, Ramachandran Ramjee2,DannyRaz1, Yuval Shavitt3, and Prasun Sinha21The Technion2Bell Laboratories3Tel Aviv UniversityDept. of Computer Science 101 Crawfords Corner Road Dept. of Electrical EngineeringHaifa 32000, Israel Holmdel, NJ 07733, USA Ramat Aviv 69978, Israel{psaar,danny}@cs.technion.ac.il {ramjee,prasunsinha}@bell-labs.com [email protected]— As local area wireless networks based on the IEEE802.11 standard see increasing public deployment, it is importantto ensure that access to the network by different users remainsfair. While fairness issues in 802.11 networks have been studiedbefore, this paper is the first to focus on TCP fairness in 802.11networks in the presence of both mobile senders and receivers. Inthis paper, we evaluate extensively through analysis, simulation,and experimentation the interaction between the 802.11 MACprotocol and TCP. We identify four different regions of TCPunfairness that depend on the buffer availability at the basestation, with some regions exhibiting significant unfairness ofover 10 in terms of throughput ratio between upstream anddownstream TCP flows. We also propose a simple solution thatcan be implemented at the base station above the MAC layerthat ensures that different TCP flows share the 802.11 bandwidthequitably irrespective of the buffer availability at the base station.I. INTRODUCTIONLocal area wireless networks based on the IEEE 802.11standard are becoming increasingly prevalent with a currentinstalled base of 15 million homes and offices. The focus isnow turning to deploying these networks over hot spots suchas airports, hotels, cafes, and other areas from which peoplecan have untethered public access to the Internet.As these networks see increasing public deployment, itis important for the service providers to be able to ensurethat access to the network by different users and applicationsremains equitable. Since the majority of applications in today’sInternet use TCP, we focus on the problem of TCP fairness inwireless LAN.Fairness issues in wireless LANs have been studied exten-sively [1], [2], [3], [4]. However, most of these solutions in-volve changes to the Media Access Control (MAC) layer. Thisis impractical given the wide deployment of these networks.Also, while the focus of previous work has been on ensuring aparticular QoS level for a given flow, we are interested in TCPfairness in the presence of both uploads and downloads i.e.in the presence of both mobile senders and receivers, whichhas not been considered by any prior work.Consider a typical installation of a 802.11 based wirelessnetwork where the mobile hosts access the network through abase station or access point. Since the 802.11 protocol allowsequal access to the media for all hosts, the base station andthe mobile hosts all have equal access to the medium. If themobile hosts are all senders or all receivers, then they eachhave equal share of the total available bandwidth. However,consider the case when there is one mobile sender and therest are all mobile receivers. In this case, the base station andthe mobile sender get equal access to the media. This mobilesender, therefore gets half of the channel bandwidth and theremaining half is equally shared by all the mobile receivers.Depending on the number of receivers, the sender can achieveseveral times the bandwidth of the receivers. Thus, the veryequal access nature of the 802.11 media access protocol,when applied to the standard installation of access througha base station results in significant unfairness. This unfairnessproblem is compounded further in the case of TCP because ofthe greedy closed loop control nature of TCP and is the focusof our paper.In this paper, we evaluate extensively through analysis,simulation, and experimentation the interaction between the802.11 MAC protocol and TCP. We identify four different re-gions of TCP unfairness that depend on the buffer availabilityat the base station, with some regions exhibiting significantunfairness of over 10 in terms of throughput ratio betweenupstream and downstream TCP flows. We also propose asimple solution that can be implemented at the base stationabove the MAC layer. The solution ensures that different TCPflows share the 802.11 bandwidth equitably irrespective of thebuffer availability at the base station.The rest of the paper is organized as follows. In Section II,we present the overview of the problem of TCP fairness over802.11 networks. In Section III, we present simulation resultshighlighting the four different regions of unfairness withrespect to the base station buffer availability. In Section IV, wemodel the behavior of multiple mobile TCP hosts accessingthe 802.11 network through a base station. In Section V,we discuss approaches for solving the fairness problem andpresent our solution. In Section VI, we review related workand finally in Section VII, we present our conclusions.II. PROBLEM OVERVIEWIn order to illustrate the subtle interactions of TCP with anunfair 802.11 MAC protocol, consider the simple case of onemobile TCP sender and one mobile TCP receiver interactingwith the wired network through a base station.We conducted a series of performance tests on a com-mercial 802.11b network consisting of one base station andthree mobile users. In all tests we had two or three mobilestations communicating to a server through the base station.Table I summarizes the throughput ratios we observed in thedifferent settings with Rurepresenting the average TCP uplink0-7803-7753-2/03/$17.00 (C) 2003 IEEE IEEE INFOCOM 2003MTU #ofupflows # of down flows UDP flow Ru/RdSD1500 1 1 – 1.44 0.221500 2 2 – 1.58 0.231500 3 3 – 1.76 0.341500 4 4 – 1.80 0.271500 2 2 500/2ms 1.79 0.351500 2 2 1000/2ms 2.15 0.55500 1 1 – 1.77 0.39500 2 2 – 1.83 0.38500 3 3 – 1.87 0.41500 1 1 450/1ms 3.05 0.83500 2 2 450/1ms 7.9 4.57TABLE ITHE RATIO BETWEEN THE UP AND DOWN FLOW IN USING COMMERCIAL 802.11Bthroughput and Rdrepresenting the average TCP downlinkthroughput. The ratios presented in the table are the averageof 5-10 runs and the standard deviation is presented in the lastcolumn.One can see that even for the basic case of one mobilesender (upstream flow) and one mobile receiver (downstreamflow), there is no equal sharing of the bandwidth with thesender receiving 1.44 times the receivers bandwidth. This isinteresting since one might expect a commercial