Timebased airness Impr ves erf ormance in Multirate WLANs Godfre an and John Guttag MIT Computer Science and Articial Intellig ence Labor atory godfre yt guttag csail
168K - views

Timebased airness Impr ves erf ormance in Multirate WLANs Godfre an and John Guttag MIT Computer Science and Articial Intellig ence Labor atory godfre yt guttag csail

mitedu Abstract The performance seen by indi vidual clients on wireless lo cal area netw ork WLAN is hea vily in64258uenced by the manner in which wireless channel capacity is allocated The popular MA protocol DCF Distrib uted Coordination Function u

Download Pdf

Timebased airness Impr ves erf ormance in Multirate WLANs Godfre an and John Guttag MIT Computer Science and Articial Intellig ence Labor atory godfre yt guttag csail

Download Pdf - The PPT/PDF document "Timebased airness Impr ves erf ormance i..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.

Presentation on theme: "Timebased airness Impr ves erf ormance in Multirate WLANs Godfre an and John Guttag MIT Computer Science and Articial Intellig ence Labor atory godfre yt guttag csail"— Presentation transcript:

Page 1
Time-based airness Impr ves erf ormance in Multi-rate WLANs Godfre an and John Guttag MIT Computer Science and Artificial Intellig ence Labor atory godfre yt, guttag @csail.mit.edu Abstract The performance seen by indi vidual clients on wireless lo- cal area netw ork (WLAN) is hea vily influenced by the manner in which wireless channel capacity is allocated. The popular MA protocol DCF (Distrib uted Coordination Function) used in 802.11 netw orks pro vides equal long-term transmission op- portunities to competing nodes when all nodes xperience sim- ilar channel

conditions. When similar -sized pack ets are also used, DCF leads to equal achie ed throughputs thr oughput- based fairness among contending nodes. Because of arying indoor channel conditions, the 802.11 stan- dard supports multiple data transmission rates to xploit the trade-of between data rate and bit error rate. This leads to considerable ate diver sity particularly when the netw ork is congested. Under such conditions, throughput-based airness can lead to drastically reduced aggre gate throughput. In this paper we ar gue the adv antages of time-based fairness in which each competing node

recei es an equal share of the wireless channel occupanc time. demonstrate that this no- tion of airness can lead to significant impro ements in aggre- gate performance while still guaranteeing that no node recei es orse channel access than it ould in single-rate WLAN. also describe our algorithm, TBR (T ime-based Re gulator), which runs on the AP and orks with an MA protocol to pro vide time-based airness by re gulating pack ets. Through xperiments, we sho that our practical and backw ard com- patible implementation of TBR in conjunction with an xisting implementation of DCF achie es

time-based airness. Intr oduction 802.11 is the de facto wireless netw orking standard. In typ- ical deplo yment, mobile node or station equipped with an 802.11 interf ace communicates er the air to an access point (AP) or base station that is connected to wired backbone. There are number of dif ferent 802.11 standards. or con- creteness, we focus primarily on 802.11b, the most widely used ersion of 802.11. When multiple mobile nodes wish to use the wireless channel simultaneously the channel must be apportioned in some “f air ay among them. In 802.11 net- orks, the apportioning is controlled

by DCF at the MA layer and the queuing mechanism used at the APs. or reasons we discuss later nodes connected to 802.11 WLANs transfer data at number of dif ferent rates. So, for xample, the channel capacity might ha to be apportioned between nodes transferring data at 11 Mbps and nodes trans- ferring data at Mbps. In this paper we first demonstrate that DCF and the xisting queuing schemes at the APs pro vide notion of airness that is inherently inef ficient, and then pro- pose and aluate better mechanism. The signal strength and loss rate of indoor wireless channels ary widely en

for nodes that are equidistant from access points [19 ]. When the 802.11 MA detects pack et loss (due to the absence of synchronous ac ), it continues retransmit- ting the pack et until the maximum retry limit has been reached. Ho we er this is futile when the erage signal strength at the recei er is consistently lo wer than the threshold required for successful pack et reception. In such cases, the sender can transmit at lo wer data rate (using more resilient modula- tion scheme) so that the channel bit error rate (BER) is re- duced. In general, there is trade-of between data rate and BER [11

16 ]. Man endors of APs and client cards implement automatic rate control schemes in which the sending stations adap- ti ely change the data rate based on percei ed channel condi- tions [7 16 21]. Man cards also allo users to manually set the data rate. The 802.11b standard defines four dif ferent data rates, and 11 Mbps respecti ely This leads to ate di- ver sity in the system, where competing nodes within cell use dif ferent data rates to communicate with the AP (in both up- link and do wnlink directions). As sho wn in Figure 1, arious data transmission rates were used during 90-minute

sessions of student orkshop that took place at MIT Furthermore, WLANs carry significant amounts of traf fic, and thus man APs xperience se eral congested periods. In Section, 3, we discuss the pre alence of rate di ersity in more detail. When multiple nodes are simultaneously xchanging data us- ing dif ferent data rates during congested periods, the total net- ork throughput is quite dif ferent from what one might x- pect. Figure illustrates ho the aggre gate throughput can be dramatically reduced when tw competing nodes use dif ferent data rates to upload files using TCP The

achie ed throughput of the node with the higher transmission rate is reduced by about 75 times.
Page 2
0 10 20 30 40 50 60 70 80 90 100 WS-1 WS-2 WS-3 EXP-1 Percentage of total bytes 1Mbps 2Mbps 5.5Mbps 11Mbps Figure 1: Fractions of bytes transferred at arious data rates during three 90-minute orkshop sessions (WS) and an xper iment (EXP-1). The root cause of this beha vior is the definition of airness used by DCF This ariant of the CSMA medium access protocol is designed to gi approximately equal tr ansmission opportuni- ties to each competing node. That is to say each node

will ha approximately the same number of opportunities to send data frame, irr espective of the amount of time equir ed to tr ansmit pac et When the same-sized pack ets are used and channel conditions are similar each competing node, re gardless of its data rate, achie es roughly the same throughput, as sho wn in Figure 2. Since the node transmitting at Mbps will tak se eral times longer to transmit frame than the node transmitting at 11 Mbps, the channel is being used most of the time by the slo wer node. In Figure 2, the fraction of the channel time used by the slo wer node is times as much

as that used by the aster node. Hence, the total throughput is reduced to le el much closer to what one gets when both competing nodes are slo The aster node pays penalty for competing against slo node rather than against another ast node. Aggre gate throughput is also impacted. Nai ely one might x- pect the total throughput of an 11 Mbps and Mbps chan- nel to be some where around 93 Mbps, the erage of the to- tal throughputs achie ed by pair of 11 Mbps channels 08 Mbps) and pair of Mbps channels 78 Mbps). Ho we er it is only 34 Mbps, less than half of what one might xpect. And the situation

is lik ely to become orse as the emer ging 802.11g netw orks, with maximum data rate of 54 Mbps, are deplo yed alongside relati ely slo wer 802.11b netw orks. 802.11g users may see ar less performance impro ement than xpected, thus lo wering the incenti for users to upgrade to 802.11g cards. DCF mainly af fects the channel capacity allocation in the up- link direction. The pack et scheduling mechanism at the AP 0 0.5 1 1.5 2 2.5 11vs11 11vs1 0 0.193 0.386 0.579 0.772 0.965 Throughput (Mbps) Channel Occupancy Time Fraction Node-1(Thruput) Node-2(Thruput) Node-1(ChanTime) Node-2(ChanTime) Figure

2: TCP throughputs achie ed and fractions of channel occupanc time used by tw competing nodes when i) both sending at 11 Mbps and ii) one sending at 11 Mbps and the other at Mbps. dictates the channel capacity allocation to clients in the do wn- link direction. When there are multiple backlogged pack ets destined to more than one clients, the scheduling scheme must decide the order of transmission. Again, since the channel con- ditions at the clients ary dif ferent data transmission rates are often used for dif ferent clients. Scheduling schemes in the literature [8 9, 24 pro vide

throughput-based airness that has been widely-accepted in wired netw orks and single-rate 802.11 WLAN, in which the data rate for each transmission on the shared medium is the same. When such scheduling schemes are emplo yed at the APs of multi-rate WLANs, the channel capacity allocation on the do wnlink direction is im- pacted in similar undesirable ay as in the uplink direction. belie that these inef ficiencies are best addressed by adopting notion of airness that gi es each competing client node an approximately equal amount of the shared channel resource: hannel occupancy time This

notion of of ime- based airness is quite dif ferent from the throughput-based airness notion widely accepted in wired netw orks and single- rate WLANs. ime-based airness pro vides an important prop- erty in multi-rate WLANs that throughput-based airness does not: Baseline pr operty: The long-term throughput of node com- peting against an number of nodes running at dif ferent speeds is equal to the throughput that the node ould achie in an xisting single-rate 802.11 WLAN in which all competing nodes were running at its rate. I.e., the throughput node achie es when competing against nodes is

identical to what it ould achie if it were competing against nodes all using its data rate. airness is, of course, subjecti notion (as an parent of
Page 3
multiple children kno ws). do not claim that one notion is “f airer than the other Ho we er we do point out that in the presence of rate di ersity during congested periods, time-based airness does impro the erall netw ork performance. In this paper we: Examine the impact of both time-based and throughput- based airness on arious measures of netw ork ef ficienc Present an analytic frame ork in which the impact of rate di

ersity on the netw ork performance is quantitati ely aluated for each airness notion used alidate our model against deplo yed 802.11b netw ork Sho by collecting and analyzing trace data, that cur rent 802.11b netw orks indeed suf fer the predicted perfor mance de gradation in the presence of rate di ersity Present an ef fecti and ef ficient scheme, TBR (for ime- based Re gulator), for deplo ying time-based airness in x- isting AP-based WLANs, irrespecti of the MA proto- col used Describe an ef ficient 802.11-based implementation of TBR that requires changing only the dri er on the

access point, and Demonstrate the relati adv antage of time-based air ness, both analytically (using our model) and xperimen- tally (using the 802.11-based implementation) The rest of this paper is or ganized as follo ws. Section ana- lyzes the xpected performance impact of both notions of air ness and xamines which notion of airness DCF achie es un- der arious circumstances. Section presents netw ork trace analyses and xperiments that demonstrate that rate di ersity is common in today netw orks. Section describes in detail our scheme to achie the time-based airness, Section alu- ates our

scheme performance and Section discusses related ork. Analysis In this section, we ar gue why time-based airness is desirable in some cases and analyze the achie ed throughputs of com- peting nodes, possibly using dif ferent data rates and pack et sizes, in 802.11-lik CSMA WLANs. 2.1 Impact of air ness Notions on Efficiency no xamine ho dif ferent notions of airness impact the erall ef ficienc of multi-rate WLANs. The measure of air ness between nodes and with equal priorities during an in- terv al is: where and are their achie ed portions of the shared resource. In this paper we

only focus on nodes with equal priorities. Dif- ferent notions of airness are captured by dif fering definitions of Let and be the channel occupanc time and the achie ed throughput respecti ely of node dur ing The choice of airness notion dictates ho the netw ork allo- cates the shared resource (in our case channel capacity) during periods in which demand xceeds supply The erall netw ork performance as well as the performance of indi vidual nodes can be greatly af fected by it. define netw ork ef ficienc as the sum of the utility of each competing node based on their shares

of shared resource. use tw traf fic models, fluid model [8, 27 and task model [4 ], to xamine the impact of airness notions on erall netw ork ef ficienc In the fluid model, there is finite number of flo ws, each of which continuously transfers infinite streams of bits. The net- ork ef ficienc can be aluated using its (a erage) aggre gate sustained throughput Ag grThruput ). Note that while the in- stantaneous throughput of node will ary depending upon the its data rate, the xpected instantaneous total throughput is time in ariant. In the task

model, there is finite number of flo ws, each of which transfers finite number of bits. Since we are pro viding airness only among competing nodes, we assume that each node has one flo In this model, the instantaneous aggre gate throughput aries with the remaining task mix. Thus, it is more appropriate to look at netw ork ef ficienc in other ays such as the erage task completion time, vgT askT ime and the fi- nal task completion time, inalT askT ime Short vgT askT ime is especially desirable for mobile nodes since those that ha completed their communication

tasks can turn-of their wire- less cards to sa ener gy or mo to another place to go on with their ork. Short inalT askT ime is also desirable since it implies higher long term erage aggre gate throughput and thus the netw ork can potentially accommodate more tasks. Figures 3(a) and 3(b) compare the achie ed TCP through- puts and the channel occupanc times of tw competing nodes when dif ferent airness notions (RF and TF) are used. These figures assume flo model or task model in which no flo has yet completed. The graphs are based on the xperiments we conducted. In the

remainder of this section we demonstrate that these xperimental results are consistent with analytical predictions. Observ that when both nodes transmit at the same rate (11vs11 and 1vs1), the allocations of both throughputs and channel occupanc times are identical for both airness no- tions. Ho we er when one node transmits at Mbps and the other at 11 Mbps (see middle bars in figures), nodes achie equal throughputs under throughput-based airness, ut achie es more throughput than under time-based airness. The situation is re ersed with respect to the alloca- tion of channel occupanc

time. Each node achie es an equal amount of channel occupanc time under time-based airness, ut gets much lar ger share than under throughput- based airness.
Page 4
0 1 2 3 4 5 RF TF RF TF RF TF Throughput (Mbps) n1(11) n2(11) n1(1) n2(11) n1(1) n2(1) 1vs1 1vs11 11vs11 (a) Achie ed Throughput 0 0.2 0.4 0.6 0.8 1 RF TF RF TF RF TF Channel Occupancy Time (Fraction) n1(11) n2(11) n1(1) n2(11) n1(1) n2(1) 1vs1 1vs11 11vs11 (b) Channel Occupanc ime Figure 3: Achie ed TCP throughputs and fractions of channel occupanc time of tw competing nodes in three dif ferent com- binations of data

rates: 11vs11, 1vs11 and 1vs1. thr oughput-based fairness and TF denote the throughput-based and time-based airness notions respecti ely E.g. in 3(a), n1(11) denotes the throughput achie ed by node n1 transmitting at 11 Mbps. Criteria Measure RF TF airness Better orse orse Better Ef ficienc inalT askT ime Same Same (task model) vgT askT ime orse Better Ef ficienc Ag grThruput orse Better (fluid model) able 1: Comparison of dif ferent measures when the throughput-based (RF) and time-based (TF) airness notions are enforced. Compared to throughput-based airness, time-based

airness benefits aster nodes at the xpense of slo wer nodes. Ho we er the airness property captured by the baseline pr operty of Sec- tion is maintained. Each class of node performs as it ould in single-rate 802.11 WLAN. or instance, the achie ed throughput of in both 1vs11 and 1vs1 cases is the same under time-based airness. The same statement can be made for other performance measures such as per -pack et latenc able compares arious measures of airness and ef ficienc for scenarios in which nodes within cell compete using dif- ferent data rates. As xplained in the rest of this

section, the conclusions captured in this table hold for an number of nodes. Ho we er for concreteness, we use the 1vs11 case as concrete xample. Under the task model, we assume that each node has an equal amount of data to transfer echnically the same results apply so long as each node has similar distrib u- tion of task size. When the fluid traf fic model is used, higher Ag grThruput re- sults under time-based airness as vident in Figure 3(a). inal- askT ime remains unchanged under the task model since the netw ork is ork-conserving under both airness notions. Ho w- er vgT askT

ime under time-based airness is lo wer than that under throughput-based airness. Under throughput-based air ness, vgT askT ime inalT askT ime since both tasks com- plete at the same time. Under time-based airness, in contrast, vgT askT ime inalT askT ime This is because the task of the 11 Mbps node will complete sooner since it achie es higher throughput while the completion time of the Mbps node re- mains the same. The rest of this section xamines ho well xisting 802.11 DCF achie es each notion of airness, and presents an analyt- ical frame ork to predict the netw ork performance in multi-

rate WLANs. 2.2 air ness in AP-based WLANs raditional air queuing algorithms designed for wired net- orks attempt to pro vide air allocation of the bandwidth on shared link [8 24 ]. Pre vious ork on air schedul- ing in wireless netw orks generally adopted this notion of air ness [20 22 27 ]. Ho we er unlik wired links, typical wire- less netw orks are half-duple ed in that the channel needs to be shared for both transmitting and recei ving pack ets. In AP-based WLANs, each AP is just acilitator and thus the resource used by it to transmit pack ets destined to client should be accounted as part

of the resource used by the client or its flo In the rest of this paper we focus on pro viding air channel time shares among competing nodes, not flo ws. The channel time used by competing node is the total chan-
Page 5
nel time used in both transmitting and recei ving pack ets to and from the AP belie that this notion is more intu- iti than the traditional notion of pro viding air resource al- locations among competing flo ws. The latter is more suitable for wired netw orks and ad hoc wireless netw orks, where there are no acilitators present (e.g. when the medium

is shared by nodes in distrib uted manner) or the acilitator is the only one transmitting on the medium (e.g. router scheduling pack ets to transmit on an output link). 2.3 Netw ork Model In this subsection, we describe the netw ork model that we use to analyze the performance of AP-based WLANs. In these infrastructure-based WLANs, each wireless node only com- municates directly with an AP in order to xchange data with another node inside or outside of the WLAN. As in much of the xisting literature [8, 27 ], we base our anal- ysis on the fluid traf fic model, and thus are concerned

with Ag- grThruput Ho we er the results in this section clearly indicate that when the task traf fic model is used, the netw ork ef ficienc in terms of vgT askT ime is better under time-based airness than under throughput-based airness (see Section 2.1). Let be the set of competing nodes and its cardinality define and as the data rate used and data pack et size used by node or simplicity of analysis, we assume that and apply to data pack ets in both uplink and do wnlink directions of node define the hannel occupancy time of node as the fraction of time wireless node is

able to access the channel to either transmit or recei pack ets to and from the AP The channel occupanc time necessary to trans- fer data pack et includes i) the transmission time of the data pack et, ii) the transmission time of synchronous ac iii) the propagation delays, v) the inter -frame idle periods necessary for node to be idle before accessing the channel, and v) the amount of time required to perform retransmissions when nec- essary Since we assume that the channel is usy all the time: (1) Let and be the total throughput achie ed by all nodes in and the achie ed throughput of node

respecti ely can xpress in terms of as: (2) where is the baseline thr oughput (that nodes xpe- rience) as function of the data rate, and the pack et size, holding all else equal. The baseline throughput equals the maximum total achie ed throughput when all nodes use the same pack et size and data rate under similar loss characteristics. or instance, when tw nodes simultaneously transfer files using 1500 -byte TCP pack ets and data rate of 11 Mbps, the baseline throughput (as sho wn in Figure 2) is 5.08 Mbps. Ho we er the actual throughput node de- pends upon the fraction of time as able

to access the chan- nel, The total actual throughput of the netw ork is simply: (3) Baseline throughput increases with the increase in data trans- mission rate as well as pack et size. The latter is due to re- duced per -pack et erhead as result of the lar ger number of payload bits per pack et. By xpressing in terms of we oid dealing directly with other actors that af fect the throughput such as the back-of periods and physi- cal layer erhead, that are independent of the ork co ered in this paper d; s; can be obtained both theoretically and xperimentally In Section 2.7, we report measured

alues of d; s; for arious alues of Furthermore, we do not deal with arying loss characteristics since our goal is in under standing ho di erse data rates and pack et sizes af fect the net- ork performance. 2.4 Impact of DCF on air ness Notions 802.11 DCF (Distrib uted Coordinating Function) is ar -and- ay the most commonly used contention resolution method in 802.11 netw orks. Although an alternati Point Coordinat- ing Function (PCF) xists, it is not implemented by most AP endors because of its comple xity and issues of co-e xistence with DCF-based netw orks. DCF gi es equal transmission op-

portunities (or long-term channel access probability) to each contender [17 26 ]. Therefore, competing nodes attempting to send data pack ets to the AP er the same time interv al will be able to trans- mit equal numbers of frames. DCF transmission opportunity based mechanism pro vides air allocations of both throughput and channel occupanc time only if all contending nodes i) use the same date rate, ii) use the same pack et size, and iii) x- perience ery similar loss characteristics. If only the last tw conditions hold, DCF achie es throughput-based airness ut does not achie time-based

airness. or an other combina- tion, DCF achie es neither time-based airness nor throughput- based airness. Figure sho ws the throughputs achie ed by three competing nodes that are either sending or recei ving data using the max- imum data rate of 11 Mbps and the maximum pack et size of 1500 bytes. In uplink directions, the throughput achie ed by each node is approximately equal due to DCF In do wnlink directions, the throughput achie ed by each node is approx- imately equal lar gely due to the AP queuing scheme, which usually transmits to wireless clients in round-robin manner TCP throughputs

are significantly less than UDP throughputs because the transmission erhead of TCP ac pack ets. The to- tal throughputs achie ed in the uplink direction are higher than those in the do wnlink direction. This is because one 802.11
Page 6
0 0.5 1 1.5 2 UDP_Down UDP_Up TCP_Down TCP_Up Total Achieved Throughput (Mbps) Node-1 Node-2 Node-3 Figure 4: UDP and TCP throughputs achie ed by three com- peting nodes (Cisco-350 cards) each of which is xchanging data at 11 Mbps with common AP (Cabletron Roamabout- 2000). “Up and “Do wn x-axis labels denote that the nodes are sending data to

and recei ving from the AP respecti ely sending node (the AP) cannot fully utilize or saturate the chan- nel since transmitting node is required to back-of for random period, between and 610 us, after ery success- ful pack et transmission. This erhead is reduced with the in- crease in number of competing nodes. no deri the general xpression of the fraction of time node is able to transmit or recei pack ets under DCF or ease of notation, we will use in place of or steady state performance, we can assume that in each round, each competing node transfers single pack et. Thus, is simply the ratio

of the time required for node to transfer data frame, which is to the total time required for ery node in to transfer data frame. (4) 2.4.1 Impact of Rate Diver sity understand the impact of rate di ersity let assume that each node uses the same pack et size, i.e. i; Therefore, based on Equations and 3, (5) (6) (7) Equation clearly sho ws that the throughput of each node is the same. Thus, under these conditions, DCF achie es throughput-based airness. Observ e, ho we er the amount of throughput is dependent on the baseline throughputs of all nodes in which in turn depend on their data rates

and pack et sizes. The channel occupanc time of node is in ersely pro- portional to the baseline throughput of node which increases with the increase in transmission rate. Thus, as xpected, nodes with slo wer data rates occup the channel much longer than those with higher data rates, leading to de gradation in the er all netw ork performance. 2.4.2 Impact of ac et Size Diver sity The impact of pack et size di ersity can be understood by as- suming that each node uses the same data rate, i.e i; Based on Equations and 3, we ha e: (8) (9) (10) Once again, depends on the baseline throughputs of

all other competing nodes. Ho we er the equations mak it clear that in this case and may dif fer across nodes, de- pending upon pack et size. 2.5 Impact of the AP Queuing Scheme The queuing mechanism at the AP dictates the channel band- width allocation to clients in the do wnlink direction. Since the channel conditions at the clients ary dif ferent data transmis- sion rates are often used for dif ferent clients. As ar as we kno the xisting literature on scheduling schemes [20 22 27 does not consider the impact of rate di ersity Thus, the aggre gate netw ork throughput when only do wnlink traf

fic is present is impacted in the same ay as pre viously xplained. also note that if loss rates xperienced by nodes dif fer and both pack et transmissions in uplink and do wnlink directions use dif ferent data rates, the achie ed throughputs of compet- ing clients may not be equal or easily predictable, en when all nodes use DCF and the AP emplo ys air queuing scheme. 2.6 Impact of ime-based air ness Under the time-based airness, our proposed definition of air ness, each node achie es an equal share of channel occupanc time. Thus, (11) Substituting Equation 11 in Equations and 3,

Page 7
Mbps) (Byte) d; s; 11 1500 5.189 5.5 1500 3.327 1500 1.493 1500 0.806 able 2: The xperimentally achie ed total throughput (or the baseline throughput) of the tw nodes simultaneously x- changing data at the same data rate and pack et size Each node has similar frame loss rate of less than 2% airness 1) 2) 3) 4) otal Criteria (1) (2) (11) (11) RF 0.436 0.436 0.436 0.436 1.742 TF 0.202 0.373 1.30 1.30 3.175 able 3: Comparison of achie ed throughputs (in Mbps) of four nodes, each transmitting at 1, 2, 11 and 11 Mbps respec- ti ely under RF and TF Note that 1) under TF is the

same as what ould achie if all and transmit at Mbps. (13) Notice that only depends on what node can achie un- der the gi en conditions and the number of competing nodes. It does not depend upon the data rates or pack et sizes used by competing nodes. Unlik sho wn in pre vious subsec- tions, is simple summation of each node maximum achie able throughput when all competing nodes use its data rate and pack et size. and in Equations and 10 will be equal if and only if all nodes in use the same data rate and pack et size. 2.7 Examples In this section we illustrate the ramifications of the dif

ferences between Equation 12 and Equation with small xample. able sho ws the xperimentally deri ed baseline throughputs of tw identical competing nodes as function of transmission rate. This pro vides an estimate of baseline throughput for ar ious transmission rates. Using these alues, we compute the throughputs when con- tains four competing nodes, one communicating at Mbps, one at Mbps, and at 11 Mbps. These are sho wn in able 3. The achie ed throughput of the slo wer nodes is less under time-based airness than under throughput-based airness. Un- der time-based airness, the Mbps and Mbps

nodes achie the throughput the ould ha achie ed of all four nodes were running at their speed. The 11 Mbps nodes achie con- siderably higher throughput under time-based airness, and the total throughput impro es by 82% Existence of Rate Di ersity In this section, we discuss in detail i) whether rate di ersity xists in today 802.11b netw orks and ii) whether single user or multiple users are acti ely xchanging data during the interv als in which the netw ork is saturated. in estigate the pre alence of rate di ersity we collected traces of wireless netw ork traf fic at one-day Iris student

ork- shop at MIT There were about 45 attendees and more than half turned on their wireless laptops. set up laptop to snif data during each of the three 90 -minute sessions, WS-1, WS-2 and WS-3, all of which took place in single room of about 40 25 Figure sho ws the fractions of data bytes transferred using each of the four possible rates during each session. It is clear that rate di ersity xists en in relati ely small room. During WS-2, more than 30% of the data bytes were transferred using data rates lo wer than 11 Mbps. also set up an xperiment to in estigate ho an AP change data rates to

arious clients in indoor of fice en vironments. placed Cabletron Roamabout-2000 AP in 18 14 of fice abo ground. sender with wired connection to the AP sent unicast UDP data pack ets at the saturation rate simultaneously to four dif ferent recei ers. The first node as about ay from the AP the second 12 and one thin, ooden all ay the third 26 and tw thin ooden alls ay and the fourth 30 and tw thick alls in between. As sho wn in Figure (see EXP-1), more than 50% of the bytes were transferred using the lo west data rate. In act, recent xtensi wireless netw ork usage study on uni

ersity campus has found that the erage recei ed signal strength aries widely en among positions that are within 20 of an access point [19 ]. Thus, we belie that rate di ersity is pre alent in man indoor WLANs and its impact ould be much more pronounced with mix ed deplo yments of 802.11b and 802.11g netw orks. The ne gati impact of rate di ersity is significant only if the follo wing tw conditions are true: i) more than one compet- ing node xchange data during the periods in which netw ork is saturated and ii) competing nodes use di erse data rates. Our analysis of this particular

orkshop trace data, ho we er sho ws that the netw ork is well er -pro visioned with APs pro vid- ing combined channel capacity of 33 Mbps. Ho we er recent studies ha sho wn that in man enterprise netw orks [2] and uni ersity residential halls [18 ], WLANs carry significant traf- fic and contain man APs that ha lot of usy or congested periods. analyzed wireless tcpdump trace of Whittemore, residen- tial acility in the Dartmouth usiness school where students were required to wn laptops. This data as collected by otz et al. er the Spring semester [18 and as made publicly ailable by

otz. Unfortunately the trace data does not con- tain the data transmission rate used for each frame transmis-
Page 8
0 20 40 60 80 100 0 100 200 300 400 500 600 Fraction of total throughput (%) 1-second intervals in which total throughput exceeded 4Mbps Heaviest User Figure 5: The fraction of throughput achie ed by the hea viest user at usy AP during usy 1-second interv als. sion. Nonetheless, we can identify the usy periods in which an AP is carrying close-to the maximum amount of data, and in estigate whether more than one user acti ely xchange data during congested periods. Since

TCP dominated the traf fic, we conserv ati ely define usy or cong ested intervals as those in which the total data through- put at the AP xceeded Mbps, 80% of the commonly ob- serv ed TCP saturation throughput when nodes transmit at the maximum data rate and xperience ery lo loss rate of 1% to 2% Figure plots the fraction of aggre gate throughput achie ed during usy 1-second interv als by the hea viest user at an AP at Whittemore on April 2002, Spring Monday The hea viest user is one that xchanged the most bytes with the AP Al- though the majority of bytes were transferred by one

user on erage, it is clear that the hea viest user alone rarely saturated the channel. In most 1-second usy interv als, users other than the hea viest user xchanged significant amounts of data. ime-based Regulator In the pre vious sections, we ha ar gued that competing nodes should be gi en an equal amount of long-term channel occu- panc time. As xplained before, in AP-based WLANs, the MA protocol and the queuing scheme at the AP in com- bination determine the channel time allocation. Therefore, to achie desired channel time allocation, coordination is nec- essary between the MA protocol

and the queuing scheme. Our proposed ime-based Re gulator runs at each AP coordi- nates with clients when necessary and orks in conjunction with an MA protocol. TBR pro vides an equal share of long-term channel occupanc time to each competing client node by Dictating ho pack et transmissions are scheduled at the AP as well as at the clients tok ens init buck et init ate air share of channel occupanc time initialize ueue for each buck et tok ens tok ens ate if tok ens buck et tok ens buck et destination of enqueue to ueue for each station starting with nexti if ueue is not empty and tok ens

dequeue pack et from ueue ask the MA to transmit nexti ne xt station after channel occupanc time of if as sent by AP destination of else source of tok ens tok ens if actual star current time actual actual Figure 6: Pseudo-code of TBR aking into account the channel occupanc time of traf fic in both do wnlink and uplink directions, and aking into account arying traf fic conditions, loss rates, data rates, and frame sizes typical implementation of TBR requires no modification to the underlying MA protocol and to the dri ers of mobile clients, allo wing incremental deplo yment

and preserving back- ard compatibility Modifications to the clients, ho we er are necessary to preserv correctness in cases where the uplink UDP flo ws mak up significant fraction of the WLAN traf- fic. will discuss more on this issue in the ne xt subsection. TBR is based on the leak uck et scheme [3 ]. The fundamental unit or tok en used in the implementation is the channel occu- panc time in terms of micro-seconds. TBR only schedules the
Page 9
transmission of pack et destined to or originated from client only if the node has not used up all its ailable

channel time. Figure sho ws the pseudo-code of TBR that runs on the AP TBR sits abo the MA layer and belo the netw ork layer and is implemented in ent handlers, each of which is triggered by the upper layer timer or the MA layer When node associates with the AP (i.e. joins the netw ork), is triggered. The procedure i) creates out- put queue ueue and ii) initializes tok ens the ailable to- ens buck et the maximum amount of tok ens that the node can accumulate, and ate the rate at which tok ens are being re-filled. Whene er the upper layer has pack et to transmit, it calls TBR simply

enqueues the pack et to ueue where is the destination of TBR adjusts tok ens according to the channel occupanc time of transmitted frames originated from or destined to node Section 4.2 described ho TBR computes the channel occu- panc time. buck et determines the maximum length of the urst period in which node can transmit successi ely (if no other nodes can transmit). buck et can af fect the short-term airness and we discuss this issue later in Section 4.5. TBR sets up timer that periodically calls which for each node updates tok ens according to ate and the time elapsed since the last time

as called. ate is the rate at which tok ens are being re-filled. note that =1 ate where is the number of acti client nodes. In general, ate can ary among client nodes depend- ing on the desired airness polic If each competing node should recei an equal share of the channel occupanc time, ate Ho we er in practice, not all nodes can consume their ailable channel time according to the allocation. TBR ensures that the system remain ork conserving by adjusting the tok en rates appropriately as discussed in Section 4.3. 4.1 Scheduling Frame ransmissions Whene er the MA layer is ready to accept

ne pack et for transmission, it calls TBR decides which back- logged pack et to release as follo ws. TBR chooses one out- put queue among all the output queues with positi ailable channel time (tok ens) and dequeues pack et for transmission. The manner in which the output queue is chosen has no impact on the erall correctness since only the queues with positi tok ens are considered. Nonetheless, the order could impact the short-term airness. or simplicity and to alle viate short-term unf airness, TBR chooses the output queue among those with positi tok ens in round-robin manner note that

short- term unf airness can further be reduced by choosing the queue which has the pack et with the shortest potential final comple- tion time as in traditional air queuing schemes [8 24 ]. Once the output queue is chosen, TBR can decide which frame in the queue gets transmitted. or TCP in-order pack et deli v- ery is desirable and thus first-in-first-out discipline is prefer able. Ho we er if there are time-sensiti pack ets (used by real-time protocols), the should ha priority er TCP pack- ets with earlier arri al times. The correctness of TBR does not depend on ho pack et

to dequeue is chosen. also note that TBR orks with an uf fering scheme (e.g. RED, drop- tail), whose goal is to decides which pack ets to drop when the queue is getting full. Note that we distinguish uf fering schemes from pack et scheduling schemes. The former is re- sponsible for deciding which pack ets to drop whereas the latter decides which pack et gets transmitted [8 ]. TBR also dictates the scheduling of pack et transmissions at the clients. Specifically whene er tok ens TBR needs to x- plicitly inform node to delay transmission for short amount of time. This can be accomplished

in tw major ays. First, the TBR agent at the AP informs the client by either sending an xplicit notification pack et or piggyback such information in do wnlink pack et when possible. Second, the client mon- itors the total channel occupanc time of pack ets transmitted and recei ed and transmits only if there is ailable channel time allocated for the node. do so, the client only needs to kno ate Ho we er as we xplain in Section 4.3, TBR at the AP may update ate depending on the erall traf fic con- ditions and when that happens, TBR needs to inform the client. In both cases, client

agent is necessary at each client to com- municate with TBR at the AP choose the first method for simplicity The actual amount of communication erhead depends on the MA protocol used. TBR requires single bit in the MA header of data frame transmission to inform the client to de- lay its transmission for pre-determined amount of time. In cases where there is only uplink traf fic, TBR can still use the same procedure if the underlying MA protocol (e.g. DCF) emplo ys stop-and-go retransmission strate gy stop-and- go protocol requires the node recei ving data frame to re- ply with

synchronous ackno wledgment, which can carry the TBR notification bit. Furthermore, if the underlying MA pro- tocol emplo ys polling mechanism (such as 802.11 PCF), no xplicit communication is necessary since TBR can dictate which node gets polled. Cooperation from each client is only necessary if the client has uplink UDP flo ws that represent significant fraction of its traf fic. Studies of WLAN traf fic at uni ersity cam- puses [18 25 and at multi-day conference [1 sho that TCP accounted for more than 90% of bytes xchanged er the WLANs. TCP data pack ets are

paced by TCP ac pack ets (“ack clocking [13 ]) sent out by the recei er In typical sce- nario, all TCP data and ac pack ets go through the same AP Therefore, delaying TCP data ac pack ets at the AP has the ef fect of slo wing do wn the sending rates of do wnlink (uplink) TCP flo ws. note that our current TBR implementation does not contain the client-side implementation of TBR. As we demonstrate in Section 5, TBR without the client cooperation can ef fecti ely
Page 10
pro vide long-term channel time guarantees for TCP flo ws in both directions as well as do wnlink UDP

flo ws. 4.2 Computing Channel Occupancy ime Whene er the MA layer has either finished sending or re- cei ed pack et it triggers This procedure subtracts the channel occupanc time of from the tok ens as- sociated with node that is the source or destination of It also modifies actual the actual tok ens used since star will xplain ho TBR uses actual in the ne xt subsection. no describe ho to compute the channel occupanc time for pack et define pac et tr ansfer time as the total time required to transfer data pack et at the 802.11 MA layer which is typically the sum of i)

the transmission time of the data pack et, ii) the transmission time of synchronous MA C- layer ac when necessary iii) propagation delays for both the data and ac pack ets, and v) the inter -frame idle periods nec- essary for the sending node to be idle before accessing the channel. Since the MA C-layer may perform retransmissions upon transmission ailure, the channel occupanc time is the sum of the pack et transfer time of each transmission until has successfully been transmitted or dropped as result of an undeli erable ailure. Therefore, ailed pack ets also contrib ute to the channel

occupanc time of the sending node. aking into account retransmissions is straight forw ard in the do wnlink direction. Ho we er in the uplink direction, the AP is not are of the xact number of retransmission attempts made by the client stations. Ideally the underlying MA protocol should include retry sequence number field (about bits) in the header to indicate ho man retransmissions precede the current pack et transmission. When retransmission information is not ailable for each pack et recei ed and the necessary header modification is not an option, the AP needs to estimate the

information necessary to compute the channel occupanc time. distinguish tw types of losses at the AP: one detected at the MA layer (due to the CRC check ailure) and the other at the physical layer In the former it is highly lik ely that the MA header whose size is relati ely much smaller than the typical payload size, is not corrupted and thus the AP can determine the source address of the ailed transmission as well as the transmission rate. note that the MA layer header can be made rob ust against channel errors by transmitting at lo wer data rate. Ho we er if the frame loss is detected at

the physical layer TBR can be are of the loss ut may not kno the necessary transmission information. belie that heuristics can be de- eloped to estimate the transmission information of each loss detected at the physical layer based on i) the number of acti clients in the last fe dozen milliseconds, ii) the lik elihood of each client contending, and iii) their steady state loss rates at the do wnlink direction. plan to de elop such heuristics in the future. 4.3 eeping Channel Utilization High When traf fic contains mixture of TCP and UDP flo ws that ha arious sending rates (and

bottleneck link bandwidth), it is important to correctly determine the amount of channel oc- cupanc time made ailable to each node. Specifically TBR needs to adjust ate to reflect changing traf fic conditions. or instance, the system will be under -utilized if we gi each node of the ailable channel time ut some nodes cannot con- sume all of their ailable time shares whereas others can con- sume more if allo wed. TBR periodically adjusts ate associated with each node so that the channel utilization is ept at maximum without violat- ing the max-min airness constraint [6 14 ].

That is, the small- est ate in the netw ork must be as lar ge as possible. Subject to this constraint, the second smallest tok en rate must also be as lar ge as possible. note that DCF in conjunction with simple round-robin queuing scheme at the AP generally achie es the max-min no- tion of airness when only TCP flo ws are in olv ed. Assume that there are uplink TCP flo ws and that one flo can only consume of the channel bandwidth (the wireless hop is not its bottleneck link). DCF will allo each of the remaining flo ws to consume of channel bandwidth pro vided that the

bottleneck link of both flo ws is the wireless link. TBR with an MA protocol achie es the same airness crite- ria pro vided that the MA layer has the ork conserving prop- erty that DCF does, i.e. each client node with data to transmit contends for channel access opportunistically Notice that the max-min airness criteria does not require that the actual de- mand of each node is kno wn. Rather one can simply achie the airness goal by incrementally gi ving more channel time to each competing node that can consume all the channel time made ailable to it [3 ]. implement this general idea in

TBR. Initially each competing node starts with the desired tok en rate of TBR schedules timer ent called that periodically adjusts the tok en ate ailable to each node. As sho wn in Figure 7, com- putes the xcess capacity of the under -utilized nodes each with the actual tok en rate actual lo wer than the assigned rate by the threshold th It then computes the xcess capacity min to redistrib ute equally among nodes that ha fully utilized the pro visioned bandwidth in the pre vious round. The actual method of computing min is of little importance for the long-term correctness so long as min is

not too big. Ho we er min does af fect the responsi eness of TBR to changing traf fic conditions. will discuss more about this in Section 4.5. If min is too lar ge, the instantaneous through- puts xperienced by flo ws can significantly ary Such beha v- iors may increase the uf fer requirements at the nodes to oid TCP ac compression that can lead to pack et drops. Figure sho ws particular ay of choosing min pick,
Page 11
() for each node excess ate actual now star if excess th if excess min min if excess max else add to set min min for each node ate ate min ate ate

min for each node actual Figure 7: Pseudo-code of the tok en rate adjustment ent among all under -utilized nodes, node with the maximal x- cess capacity (the lar gest dif ference in actual and assigned to- en rate). Half of min is subtracted from tok en rate and the other half redistrib uted among nodes that ha consumed tok ens at rates close to their assigned rates. In Section 5, we sho that TBR is able to eep the channel utilization high in the presence of arying traf fic conditions. 4.4 An 802.11-based Implementation implemented TBR in the HostAP [15 dri er running on Linux PC as

proof of concept. The HostAP dri er imple- ments access point functionality so that PCs equipped with popular Prism chipset based 802.11 cards can act as APs. use unique 6-byte MA addresses as node identifiers. note that TBR requires APs to set up per -node output queue. Ho we er the total uf fer space requirement is compa- rable between normal AP and an AP with TBR. or instance, if an xisting AP has the total queue size of pack ets than TBR-equipped AP can setup queues each with pack ets, where is the number of competing nodes. or ease of im- plementation, our TBR implementation uses

FIFO queues. As xplained before, TBR can ork with an uf fering scheme. Finally we note that the current implementation of TBR does not use the retransmission information in computing the pack et transfer time ut we plan to do so in the future. Thus, TBR in some cases can cause slight biases in granting channel occu- panc time to competing nodes. Nonetheless, as we sho in Section 5, it does well in achie ving its goal. 4.5 Discussion TBR is currently intended for ensuring that each competing node recei es an equal share of channel occupanc time based on max-min airness er the long run. As we

later demon- strate in Section 5, TBR orks well when competing flo ws last for hundreds of pack ets. Although we belie that long-li ed flo ws (e.g. file transfer ap- plications) are usually the cause of congestion in enterprise and uni ersity netw orks, we ackno wledge that congestion in hot- spot access netw orks may be caused by man short-li ed flo ws with di erse data rates, each sending only dozens of pack ets. Responsi eness of TBR relies on ho it adjusts the tok en rate assigned to each competing node and ho often (see ). Furthermore, the urst period buck et in

which node can transmit successi ely also influences the re- sponsi eness of TBR as well as short-term airness. Special at- tention must be paid to pack et-le el interaction between TBR and the underlying MA so that TBR can respond to arying traf fic conditions in the order of tens of pack et transfer time. In the future, we plan to understand each of these issues in detail and mak TBR responsi for ery short-li ed flo ws as well. Lar ge buck et can xacerbate the short-term unf airness, i.e. some competing nodes do not achie their desired air shares within ery short interv al,

commonly found in 802.11 WLANs [17 ]. Short-term unf airness in its most se ere form leads to TCP ack compression in which multiple TCP acks ar ri at the sender which then sends se eral TCP pack ets suc- cessi ely leading to undesirable pack et drops at the bottleneck queue. Ho we er the TCP ack compression problem can be ef- fecti ely solv ed by pacing TCP pack ets [5 ]. TBR can potentially be modified to pro vide each competing node with the desired share of channel occupanc time (not necessarily equal). Therefore, QoS mechanisms may use TBR to pro vide QoS at xisting AP-based WLANs.

also note that although the current implementation of TBR allocates channel time to nodes, it can be xtended to allocate channel time among arious flo ws of each node. note that the 802.11e standard [12 currently being drafted defines quality of service support for the 802.11 MA C. Us- ing 802.11e, competing nodes acquire ransmission Opportu- nities TXOP ), each of which is defined as an interv al of time when station has the right to initiate transmissions. TXOPs are allocated via contention or granted through the centralized coordinator lik the AP 802.11e dif ferentiates

the probability of channel access based on the traf fic cate gories. TBR can be inte grated with 802.11e by choosing appropriate traf fic cate- gories for each competing node according to their air share of channel occupanc time. Ev aluation setup xperiments to aluate the correctness and per formance of TBR. used PIII-700MHz Linux laptop
Page 12
0 1 2 3 4 5 Exp-Normal Exp-TBR Exp-Normal Exp-TBR Exp-Normal Exp-TBR Exp-Normal Exp-TBR Total Achieved Throughput (Mbps) n1(11) n2(11) n1(5.5) n2(5.5) n1(2) n2(2) n1(1) n2(1) 1vs1 2vs2 5.5vs5.5 11vs11 (a) do wnlink 0 1 2 3 4 5

Exp-Normal Exp-TBR Exp-Normal Exp-TBR Exp-Normal Exp-TBR Exp-Normal Exp-TBR Total Achieved Throughput (Mbps) n1(11) n2(11) n1(5.5) n2(5.5) n1(2) n2(2) n1(1) n2(1) 1vs1 2vs2 5.5vs5.5 11vs11 (b) uplink Figure 8: TCP throughputs achie ed in either uplink or do wnlink direction by tw competing nodes using the same data rate. Exp-Normal and Exp-TBR denote the xperiments that were run with the AP equipped without or with TBR respecti ely n1(11) denotes the throughput achie ed by node n1 transmitting at 11 Mbps. 0 1 2 3 4 5 Eq6 Exp-Normal Exp-TBR Eq12 Eq6 Exp-Normal Exp-TBR Eq12 Eq6 Exp-Normal

Exp-TBR Eq12 Total Achieved Throughput (Mbps) n1(5.5) n2(11) n1(2) n2(11) n1(1) n2(11) 1vs11 2vs11 5.5vs11 (a) Do wn-link 0 1 2 3 4 5 Eq6 Exp-Normal Exp-TBR Eq12 Eq6 Exp-Normal Exp-TBR Eq12 Eq6 Exp-Normal Exp-TBR Eq12 Total Achieved Throughput (Mbps) n1(5.5) n2(11) n1(2) n2(11) n1(1) n2(11) 1vs11 2vs11 5.5vs11 (b) Up-link Figure 9: TCP throughputs achie ed in either up-link or do wn-link direction by tw competing nodes using dif ferent data rates. Exp-Normal and Exp-TBR denote the xperiments that were run with the AP equipped without or with TBR respecti ely Eq6 and Eq12 represent the achie ed

throughputs according to Equation and Equation 12 respecti ely n1(11) denotes the throughput achie ed by node n1 transmitting at 11Mbps. equipped with D-Link WL-650 card running the Hostap dri er as the AP and IP Qs equipped with Cisco-350 cards as competing nodes. or each type of xperiment, we ran in tw dif ferent AP con- figurations: one with TBR, Exp-TBR and one without, Exp- Normal Each data point is an erage of to 10 runs and in each run, each contending node sends about 2000 1500-byte pack ets. All throughputs measured are achie ed TCP through- puts. When the AP is run under the

normal configuration, no queue is set up in the dri er Instead, the ernel interf ace queue (with the maximum size of 110) is used to store pack ets. When the AP is run with TBR, queues each with the maximum queue size of 100 is set up inside the dri er The ernel interf ace queue is then set to 10. Thus, the total uf fer space ailable to each scheme is the same.
Page 13
Figure compares the throughputs achie ed by tw competing nodes when the AP is configured with or without TBR. When competing nodes use the same data rate, Exp-TBR and Exp- Normal yield almost identical

results, sho wing that TBR incurs little erhead. When nodes use dif ferent data rates, the throughput achie ed by each competing node as well as the total throughput dif fer significantly depending upon whether TBR is used or not. As sho wn in Figure 9(a), when TBR is used, the total achie ed throughput in the do wn-link direction increases by about 6% in the 5.5vs11 case, 35% in the 2vs11 case and 103% in the 1vs11 case. Analytical Eq6 and xperimental Exp-Normal alues agree for all the cases when the AP is configured without TBR. Similarly Exp-TBR and Eq12 sho ery similar results,

af- firming that our re gulator achie es the objecti of pro viding long-term equal channel occupanc time to competing nodes. The slight dif ferences in performance between Exp-TBR and Eq12 is due to the act that TBR needs to estimate channel occupanc time without the retransmission information ail- able. Whene er pack et loss is xperienced by node, the channel occupanc time of that node needs to be decreased accordingly ithout the retransmission information, TBR in this case slightly biased the node sending at lo wer data rate, thus decreasing the total throughput by small amount com-

pared to Eq12 In the future, we plan to xtract (from the card firmw are) or estimate retransmission information as suggested in Section 4. Figure 9(b) sho ws similar impro ements achie ed by TBR in the up-link direction. also ran xperiments in olving mix ed up-link and do wn-link TCP flo ws and found similar results (not sho wn here). Throughput Exp-Normal Exp-TBR n1 2.9434 2.9542 n2 2.1276 2.1193 otal 5.071 5.061 able 4: Comparison of achie ed TCP throughputs under Exp- Normal and Exp-TBR. Node xperienced the bottleneck bandwidth of Mbps whereas node could send as ast as it could

(TCP permitted). Both nodes transmitted at 11 Mbps. understand ho well TBR orks when traf fic contains flo ws with arious demands, we set up scenario that in olv ed tw nodes, n1 and n2 each sending TCP pack ets at the same data rate of 11 Mbps ut xperienced dif ferent bottleneck link capacities. n2 xperienced the bottleneck bandwidth of Mbps while the wireless link is n2 bottleneck. achie ed this by limiting the sending rate of the application generating TCP pack ets at n2 The xpected DCF beha vior is to gi n2 Mbps of channel bandwidth and n1 the remainder a- ble sho ws the

throughputs achie ed under Exp-TBR and Exp- Normal There is no significant dif ference between the tw sets of results sho wing that the rate adjustment algorithm described in Section 4.3 orks. Related ork note that the general idea of temporal sharing in the conte xt of multi-rate WLANs has been mentioned before by Sade ghi et al. [23 ]. The ha proposed an opportunistic rate adaptation scheme (called AR) that achie es significant throughput gain er pre viously proposed rate adaptation schemes [11 16 ]. The idea behind AR is to allo nodes that ha high- quality channel condition to

transmit more than one pack et at time taking adv antage of time-correlated channel conditions. AR simply allo ws node that can transmit at 11 Mbps times more opportunities than the node transmitting at Mbps. AR justifies this by saying that nodes are achie ving simi- lar time-shares as when the both are transmitting at Mbps. AR is DCF-based protocol mainly intended for ad hoc net- orks and requires modifications to DCF Unlik AP-based netw orks, ad hoc netw orks, in which nodes communicate with each other without using access points, are more suitable when communications among

wireless nodes are dominant or no wired infrastructure xists. In contrast, AP-based netw orks are designed for communications among wireless nodes and other nodes that can be reached via wired infrastructure to which APs are connected. Unlik the pre vious ork, we in estigate and xplain the dif- fering impacts of the airness notions on the netw ork perfor mance and our ork focuses on AP-based 802.11 netw orks in which the queuing scheme at the AP significantly impacts the channel capacity allocation. Recently Heusse et al. ha sho wn through simulations and xperiments that performance de

gradation occurs when tw nodes are sending at dif ferent data rates [10 ]. Through anal- ysis, authors sho that the node sending at lo wer data rate will achie the same throughput as other nodes sending at higher data rate. The authors do not suggest an mechanism to mitigate this ef fects. Ef forts ha been made in de eloping distrib uted air schedul- ing algorithms that are suitable for the shared wireless medium. [20 22 27 ]. Lik the schemes proposed in wired netw orks [8 24 ], these wireless scheduling algorithms [20 22 27 neither tak into account the impact of transmission rate di ersity

nor the channel resource for both do wnlink and uplink traf fic as most schemes [22 27 were tar geted for ad hoc wireless netw orks. Summary and Conclusion started by sho wing that, in the presence of rate di er sity the throughput-based airness notion implemented by the 802.11 popular MA protocol and the traditional queuing schemes at the APs leads to situation in which the aggre gate throughput is determined lar gely by the slo west node.
Page 14
ne xt presented time-based notion of airness that pro- vides an equal amount of long-term channel occupanc time to each competing

node. This pre ents aster nodes from being dragged do wn by slo wer ones. Moreo er it satisfies what we called the baseline pr operty i.e., the achie ed throughput of an competing node in multi-rate WLAN is equal to what it ould achie in single-rate WLAN in which all competing nodes transmit at its data rate. In the presence of rate di ersity using this definition of airness can lead to astly impro ed ag- gre gate netw ork throughput, more than 100% in some realistic scenarios. ne xt described practical scheme called TBR that orks in conjunction with an MA protocol to pro vide

long-term time-based airness in AP-based WLANs by appropriately scheduling pack et transmissions. sho wed that TBR can be implemented in an AP dri er in ay that is backw ards com- patible with xisting 802.11 standard. implemented our scheme in the Linux Hostap dri er running on PC used as the AP and aluated it through series of xperiments. In the ab- sence of rate di ersity the performance of our implementation is equi alent to the standard implementation. In the presence of rate di ersity it achie es the predicted gains. In today AP-based 802.11b WLANs, rate di ersity is already common as our

trace analyses sho As ne wer standards such as 802.11g are deplo yed, the problem will become orse. or an xtended period of time 802.11 WLANs will run in mix ed mode, and if 802.11g clients are slo wed do wn to run at the rate of 802.11b clients, there will be little incenti to upgrade. belie that switching to time-based airness is good option. Refer ences [1] A. Balachandran, G. M. oelk er Bahl, and Rangan. Characterizing user beha vior and netw ork performance in public wireless LAN. CM Press, June 2002. [2] M. Balazinska and Castro. Characterizing mobility and net- ork usage in corporate

wireless local-area netw ork. In Pr oc. of CM MOBISYS’03 May 2003. [3] D. Bertsekas and R. Gallager Data Networks Prentice Hall, second edition, 1992. [4] J. Bruno, E. G. Cof fman, and R. Sethi. Scheduling independent tasks to reduce mean ˛nishing time. Communications of the CM 17:382–387, Jul 1974. [5] M. C. Chan and R. Ramjee. TCP/IP performance er 3g wire- less links with rate and delay ariation. In Pr oc. of CM MO- BICOM’02 pages 71–82, 2002. [6] D.-M. Chiu and R. Jain. Analysis of the Increase/Decrease Algorithms for Congestion oidance in Computer Netw orks. Computer Networks and

ISDN Systems 17(1):1–14, June 1989. [7] Data Sheet of Cisco Aironet 350 Series Access Points. http://www.cisco.com/warp/public/cc/pd/ witc/ao350ap/prodlit/carto_in.htm [8] A. Demers, S. esha and S. Shenk er Analysis and Simulation of air Queueing Algorithm. Internetworking: Resear And Experience 1:3–26, April 1990. [9] Go yal, H. M. in, and H. Cheng. Start-time air queueing: scheduling algorithm for inte grated services pack switching netw orks. IEEE/A CM ansactions on Networking oct 1997. [10] M. Heusse, Rousseau, G. Ber ger -Sabbatel, and A. Duda. Per formance anomaly of 802.11b In Pr oc. of

IEEE INFOCOM’03 April 2003. [11] G. Holland, N. H. aidya, and Bahl. rate-adapti MA protocol for multi-hop wireless netw orks. In Pr oc. of CM MO- BICOM’01 pages 236–251, 2001. [12] IEEE 802.11 orking Group. Draft Supplement to Interna- tional Standard for Information Exchange between systems LAN/MAN Speci˛c Requirements, No 2001. [13] Jacobson. Congestion oidance and control. CM Computer Communication Re vie 18, 4:314–329, 1988. [14] R. Jain, D.-M. Chiu, and Ha we. Quantitati Measure of airness and Discrimination for Resource Allocation in Shared Computer System. echnical Report 301,

Digital Equipment Corporation, Sept. 1984. [15] Jouni Malinen. Host AP dri er for Intersil Prism2/2.5/3. http: //hostap.epitest.fi 2003. ersion 0.0.1. [16] A. Kamerman and L. Monteban. elan ii: high- performance wireless lan for the unlicensed band. Bell Labs ec hnical ournal pages 118–133, Summer 1997. [17] C. E. oksal, H. I. Kassab, and H. Balakrishnan. An analysis of short-term airness in wireless media access protocols. In Pr oc. of CM SIGMETRICS’00 June 2000. [18] D. otz and K. Essien. Analysis of campus-wide wireless netw ork. In Pr oc. of CM MOBICOM’02 CM Press, Sept. 2002. [19] D. otz,

C. Ne wport, and C. Elliott. The mistak en axioms of wireless-netw ork research. echnical Report TR2003-467, Dept. of Computer Science, Dartmouth Colle ge, July 2003. [20] S. Lu, Bhar gha an, and R. Srikant. air scheduling in wire- less pack et netw orks. IEEE/A CM ansactions on Networking 7(4):473–489, 1999. [21] ORiNOCO AS-2000 System Release Note. http://www. michiganwireless.org/tools/Lucent/ORiNO CO/ AS- 2000_Rel2_1/AS2000_R2_10_01_Readme.tx [22] Ramanathan and Agra al. Adapting pack et air queue- ing algorithms to wireless netw orks. In Pr oc. of CM MOBI- COM’98 pages 1–9, 1998. [23] B.

Sade ghi, Kanodia, A. Sabharw al, and E. Knightly Oppor tunistic media access for multirate ad hoc netw orks. In Pr oc. of CM MOBICOM’02 sept 2002. [24] M. Shreedhar and G. ar ghese. Ef ˛cient air Queuing using De˛cit Round Robin. In Pr oc. of CM SIGCOMM’95 August 1995. [25] D. ang and M. Bak er Analysis of metropolitan-area wireless netw ork. ir eless Networks 8(2/3):107–120, 2002. [26] ay and K. Chua. capacity analysis for the IEEE 802.11 MA protocol. CM/Baltzer ir eless Networks 7(2):159–171, Mar 2001. [27] N. H. aidya, Bahl, and S. Gupta. Distrib uted air scheduling in wireless

LAN. In Pr oc. of CM MOBICOM’00 pages 167 178, 2000.