Achieving Fairness in Wireless LANs by Enhanced IEEE DCF *

Transcription

1 Achievig Fairess i Wireless LANs by Ehaced IEEE DCF * Nagesh S. P. Nadiraju, Hrishikesh Gossai, Dave Cavalcati, Kaushik R. Chowdhury, Dharma P. Agrawal Ceter for Distributed ad Mobile Computig, Dept. of ECECS, Uiversity of Ciciati - Ciciati,OH (adirs, cavalcdt, kaushir, dpa)@ececs.uc.edu Mesh Research ad Developmet Group Motorola, Ic., Maitlad, Florida Hrishikesh.Gossai@motorola.com Abstract Over the past few years, Wireless Local Area Networks (WLANs) have gaied a icreased attetio ad a large umber of WLANs are beig deployed i uiversities, compaies, airports etc. Majority of the IEEE based WLANs employ Distributed Coordiatio Fuctio (DCF) i Wireless Access Poits (AP) to arbitrate the wireless chael amog Wireless Statios (STAs). However, DCF poses serious ufairess problem betwee uplik ad dowlik flows. To overcome this ufairess problem, we propose a simple ehacemet to the IEEE DCF which provides priority to the AP ad thus eables it to acquire a larger share of the chael whe required. We have demostrated the ufairess problem through systematic measuremets i a experimetal test bed of WLAN usig the legacy DCF. We also developed aalytical models to calculate the throughput of AP ad the STAs ad verify these results through thorough simulatios i s-2. We observe that our simulatio results fid i good agreemet with our aalytical models. Results show that our proposed ehacemet achieves a fair distributio of badwidth ad improves the throughput (by early 300%) for the dowlik flows as compared to the DCF, without severely affectig the performace of uplik flows. Keywords: Fairess, MAC protocols, Performace Evaluatio, Test beds, WLANs. 1. INTRODUCTION I the last decade, Wireless LANs experieced a proliferatig growth due to their flexibility ad ubiquitous ature. I particular, WLAN hotspots are typically foud i uiversities, compaies, airports, shoppig malls, etc. With the icreasig iterest i the itegratio of various wireless etworks (4G etworks), WLANs are gaiig more popularity tha ever. The capacity of WLANs has rapidly icreased from 2Mbps to 54Mbps ad proposals (IEEE ) to achieve early 100Mbps are also uderway. May efforts are also beig made to make QoS provisios for real time traffic (IEEE e). Most of the curret WLAN implemetatios are based o the IEEE [1] stadard, which supports two basic mechaisms for chael arbitratio: Distributed Coordiatio Fuctio (DCF) ad Poit Coordiatio Fuctio (PCF). The implemetatio of DCF i IEEE compliat devices is madatory while provisio of PCF is optioal. DCF is based o the traditioal CSMA/CA paradigm ad provides equal chael access privileges to all participatig Wireless Statios (STAs). I cotrast, PCF is a cetralized schedulig algorithm. It requires a poit coordiator (PC) at the AP to cotrol the chael access. The default schedulig algorithm of IEEE PCF is a roud robi scheme ad may ot always be ideal. Due to the iheret complexity ivolved with the deploymet of PCF [2], most of the curret implemetatios of IEEE , eve i hot spot scearios, use DCF access mechaism. However, DCF poses serious ufairess problem betwee uplik ad dowlik flows. With DCF, the chael share of the AP would be a fractio of total umber of trasmittig STAs i its service area. All STAs, icludig the AP, have the same chael access privileges. As a result, the share of the chael obtaied by the AP is early equal to the share of ay other STA uder its coverage. This results i ufair sharig of the badwidth amog uplik ad dowlik flows. All the dowlik flows (flows that are destied for wireless statios) have to utilize the AP s chael share while the uplik flows origiatig from differet STAs ejoy a larger share. With the icrease i the umber of STAs uder the AP s coverage, the dowlik flows would suffer from relatively low share of the available badwidth. STA 1 STA 2 W1 S AP Primergy STA 5 Primergy STA 6 STA10 STA 9 Figure 1: IEEE DCF based WLAN Access Sceario I order to overcome this ufairess problem, we propose a simple MAC layer ehacemet to IEEE DCF, called Bidirectioal-DCF (BDCF). We specifically address the uplik/dowlik ufairess by providig the AP with more cotetio free trasmissio opportuities whe high *This work was supported by the Ohio Board of Regets Doctoral Ehacemet Fuds /06/$ IEEE

2 load is experieced. I particular, if a AP s MAC receives a DATA packet, istead of trasmittig a regular MAC layer ACK, it checks the buffer for a outstadig packet to ay of the STAs i its Basic Service Set (BSS). If a packet is foud, the it will sed the DATA with a piggybacked ACK after SIFS time, thus elimiatig the eed for a fresh chael cotetio to trasmit this packet. I this way, the AP gets a preferetial treatmet resultig i a relatively higher badwidth share as compared to its STAs. Clearly, this kid of preferetial treatmet for the AP is desirable i hot spots scearios. This is because most of the users accessig iteret i these hotspots use applicatios (e.g. , web browsig, Iteret radio, etc.) that typically geerate large volume of dowlik traffic for a sigle uplik request. It is worthwhile to ote that if the AP does ot have traffic to sed i dowlik directio, BDCF works exactly as DCF. Our specific cotributios through this paper are: Experimetal demostratio of the ufairess problem i a IEEE b based test-bed. A simple ehacemet to DCF for overcomig the ufairess problem i AP based etworks. A aalytical model to evaluate the throughputs of AP ad STAs complemeted by extesive simulatio study. Our aalytical ad simulatio results show that BDCF has a better throughput ad delay performace for dowlik flows ad also has a fair chael sharig i both directios as compared to DCF. I additio, we compare BDCF with DCF+ [5], which uses a similar idea to reduce MAC layer overhead ad icrease the throughput, while it does ot give ay priority to the AP. The outlie for the rest of this paper is as follows. I sectio 2, we experimetally demostrate the ufairess problem caused by DCF i WLAN hot spot scearios. We describe our proposed BDCF mechaism i sectio 3. Next, we develop a aalytical model for the uplik ad dowlik throughputs with BDCF i sectio 4. Sectio 5 provides comprehesive simulatio results, comparig BDCF with DCF ad DCF+. I sectio 6, we discuss the related work, ad fially, we coclude the paper i Sectio 7, highlightig some ope problems ad future research directios. 2. THE DCF UNFARINESS PROBLEM IN WLAN HOT SPOTS I this sectio we illustrate the ufairess amog dowlik/uplik flows whe DCF is employed i a WLAN. We have set up a experimetal test bed to model the typical WLAN hot spot sceario as show i Figure 1. We cofigured a ifrastructure based IEEE b etwork with a data rate of 11Mbps ad coected the AP to a desktop PC (with a Itel Petium 4 processor). We start the server S i the same machie, so that the coectio betwee the AP ad the server is ot a bottleeck. Furthermore, the AP uses the hostap driver package [12], ad the STAs use IEEE b USB adapters, Netgear WG311 or Atheros wireless cards. We have systematically examied the TCP ad UDP throughput performace with a symmetric ad a asymmetric traffic cofiguratio. TCP ad UDP traffic is geerated by Iperf v1.7 [13] which rus i both the cliet ad server modes. We first cosider a asymmetric traffic sceario where 3 UDP flows origiatig at the differet wireless odes destied to the server (called uplik flows) ad 7 UDP flows from the server towards the wireless odes (called dowlik flows). Each flow geerates traffic at a rate of 3Mbps, which is eough to saturate the wireless lik. We have coducted 10 differet rus for each traffic sceario ad measured the throughputs of idividual flows. The throughputs of flows i the same directio do ot have much variatio. Thus, for the sake of brevity, we oly show the aggregate throughputs of uplik ad dowlik flows for all the four scearios i Figure 2. As ca be see from the plot (i Figure 2 (a)), i the case of asymmetric UDP traffic, the three uplik flows obtai a throughput of Kbps (749.6 Kbps per flow o a average) while the 7 dowlik flows obtai a throughput of oly Kbps ( Kbps per flow o a average). It ca be observed that the throughput of a idividual dowlik flow is early 1/3 of the throughput achieved by ay uplik flow. Clearly the MAC level fairess achieved by DCF leads to a udesirable situatio. For example, cosider a typical WLAN hotspot sceario where a couple of studets (who are a part of peer-to-peer file sharig etwork) are uploadig sogs/movies through the wireless etwork. Their applicatios (are uplik) cosume fairly large share of the wireless chael ad thus limit the badwidth for the dowlik traffic. Thus other users who are checkig their e- mail or usig other predomiatly dowlik traffic based applicatios experiece larger dowload delays ad icreasig frustratio. These results motivate the eed for a preferetial treatmet to the AP i order to allot a fair share of badwidth for the dowlik flows. We propose a simple ehacemet to DCF i the ext sectio, which prioritizes the AP without the requiremet of ay additioal iformatio ad achieves the required fairess amog the uplik ad dowlik flows. 3. BIDIRECTIONAL DISTRIBUTED CO- ORDINATION FUNCTION (BDCF) I order to hadle the ufair badwidth availability to the dowlik traffic i WLAN hot spot scearios, we propose a Bidirectioal DCF (BDCF), which provides preferetial treatmet to the dowlik flows at the AP. For implemetig BDCF, we modified IEEE DCF to support piggybackig of ACK packets i the DATA trasmissio from the AP. Similar to IEEE DCF, BDCF supports both basic (DATA-ACK) ad 4-way hadshake (RTS-CTS- DATA-ACK) chael access mechaisms. I the remaider of this sectio we describe the details of BDCF

3 Throughput (kbps) 4000 Up Lik 3500 Dow Lik Asymmetric (3Up-7D) Symmetric (5Up-5D) Throughput (kbps) Asymmetric (3Up-7D) Up Lik Dow Lik Symmetric (5Up-5D) STA(i) DIFS SIFS RTS DATA 1 AP SIFS SIFS CTS DATA AP NAV(RTS) NAV(DATA 1) STA(j) Other (Source Rage) (Destiatio Rage) SIFS ACK NAV(RTS) NAV(DATA 1) DIFS NAV(CTS) Defer Access Cotetio Widow Backoff Starts Figure 2 (a) UDP Traffic Figure 2 (b) TCP Traffic Figure 3. Timig diagram of BDCF with RTS- CTS hadshake 3.1 BDCF with RTS-CTS hadshake We first start by describig the operatio of BDCF whe RTS-CTS hadshake is eabled. Figure 3 shows the timig diagram of the BDCF operatio. Iitially all STAs ad the AP coted for the chael with equal privilege. Durig the cotetio resolutio period, if the AP gets the access to the chael, BDCF works exactly like DCF. The differece comes whe a STA wis the chael cotetio. Figure 3 illustrates this case i detail. STA(i) iitially seds a RTS to the AP i order to reserve the chael for its DATA trasmissio. The AP respods this request of STA(i) by sedig a CTS. Upo receptio of the CTS packet, STA(i) seds its DATA packet (DATA i ) to the AP. After receivig DATA i, the AP checks its MAC buffer for ay packet to trasmit. If o packet is foud, the AP simply seds the correspodig ACK to STA(i). However, if the AP has a outstadig packet for ay of its servig STA(j) (where j may be same as i), it trasmits that DATA packet (idicated by DATA AP i Figure 3) with a piggybacked ACK, after a SIFS time period. If STA(i) seses trasmissio from the AP after a SIFS period, it will implicitly recogize the ACK set by the AP. As log as the STAs are withi the coverage regio of the AP, they are able detect the piggybacked ACK (STA(i) was waitig for a ACK for DATA i from AP). Sice all the STAs i the servig BSS should be withi the coverage of AP our assumptio is valid. After a successful receptio of the DATA AP, the destiatio STA(j) will sed back a ACK to the AP i the usual way. However, if the AP does ot receive the iitial data frame (DATA i ) correctly, it does ot sed ay DATA AP or ACK, thus STA(i) becomes aware of its usuccessful trasmissio ad schedules a retrasmissio. It is importat to ote that BDCF allows piggybackig of DATA packets oly at the AP, thus avoidig the chace of formig ay cycles (pheomeo where AP ad ay oe of STAs repeatedly access ad capture the chael). Furthermore, whe the AP trasmits the DATA packet with piggybacked ACK (DATA AP ), it freezes its backoff timer, such that the trasmissio of DATA AP is totally trasparet to the regular chael cotetio at the AP. The chages made at the MAC layer for implemetig BDCF does ot require ay chages to the upper layers ad thus totally trasparet to the upper layers. As we will show i the ext sectio, by usig this simple ehacemet, BDCF ca provide fair access to the wireless chael i both the directios, irrespective of the type of trasport layer (UDP or TCP) uder cosideratio. It should also be oted that with BDCF, we avoid sedig ACK, RTS, ad CTS packets while seakig a DATA AP packet from the AP, thus reducig MAC layer cotrol overhead as compared to DCF. Moreover we also reduce the time wasted i ay chael cotetio ad backoff mechaism. I order to avoid ay egative effect of BDCF o the uplik flows whe there are fewer dowlik flows tha uplik flows, we adopt a dyamic piggybackig strategy. With this strategy the AP records the umber of STAs trasmittig the uplik flows ad the umber of STAs receivig dowlik flows over a time widow. The AP piggybacks a DATA packet oly with a probability equal to the ratio of dowlik ad uplik flows. This way, we esure that the dowlik flows do ot get ay udue advatage ad ifluece fairess of the system. The workig priciple of BDCF i the basic access mechaism (without RTS-CTS hadshake) is similar to its operatio whe RTS-CTS hadshake is eabled. 4. ANALYTICAL EVALUATION OF BDCF As poited out i sectio 3, BDCF esures that the dowlik flows gets a fair share of the system badwidth. We validate this claim by derivig aalytical expressios for the throughput of the AP (dowlik) ad the remaiig STAs (uplik) whe BDCF is used with the RTS-CTS hadshake mechaism. We assume fixed STAs ad oe AP, each of them havig a packet to trasmit at all times (saturatio traffic coditios with equal umber of uplik ad dowlik flows). We also assume perfect chael coditios ad o hidde termials. We defie the throughput by the followig equatio E[payload iformatio trasmitted i a slot time] S = E[legthof slot time] We first study various evets that ca occur i ay arbitrary slot, the time duratios for the idividual evets, ad fially the total trasmissio time. It is assumed that each statio icludig the AP trasmits i a radomly chose slot time with probability ad collisios occur with a costat probability p, irrespective of the umber of previous collisios before a successful trasmissio. We cosider the (1)

4 slot time to be of legthσ. Let E p, H ad deote the average packet payload size, the packet header size (calculated as PHY hdr + MAC hdr ) ad the propagatio delay, respectively. E p ad H are measured i time uits. For simplicity we defie, W = CW mi ad m as the maximum backoff slots so that, CW max = 2 m * CW mi.the probability of trasmissio () i a give time slot as derived i [14] is 2(1 2 p)(1 p) τ = m (1 2 p)( W + 1) + p( W (1 2 p ( ) ) I what follows, we idetify the various evets that ca occur i a arbitrary time slot uder BDFC operatio ad calculate their probabilities of occurrece. Recall from sectio 3 that i BDCF, the AP trasmits a data packet uder two coditios: by cotetio or by piggybackig. I. Whe AP coteds for the chael tr1(ap) Oce the AP wis the chael, the time spet for a successful trasmissio of the DATA packet by the AP is calculated as: T s1(ap) = O rts + {H + E p + + SIFS + ACK + + DIFS}, (4) Where, O rts = {RTS ++ SIFS + CTS+ + SIFS} is the time required for RTS/CTS exchage. I case of a collisio, whe RTS/CTS scheme is used, the collisio time is give by T c1(ap) = RTS + + SIFS + CTS+ + DIFS (5) II. Whe the AP Piggybacks Wheever the AP receives a DATA packet from a STA, it will have the optio of sedig a DATA packet with the piggybacked ACK after SIFS time period. We deote this trasmissio as tr2. With the assumptio of saturatio traffic coditios, the AP always has a DATA packet to sed to ay oe of the STAs. As the chael is reserved for the ACK trasmissio from the AP, this trasmissio (DATA with piggybacked ACK) is guarateed (assumig o hidde termials) ad there is o eed for cotetio. Thus the collisio probability i this case is 0. The trasmissio probability for the AP (P tr2(ap) ) i this case is the same as the probability of successful trasmissio by ay STA i the prior data trasfer stage, which is give by P tr2(ap) = ) = τ I this case, the AP coteds for the chael afresh with all the other odes (this evet is deoted by tr1). I this sceario, BDCF works similar to DCF ad collisios ca occur whe two or more RTS packets are trasmitted simultaeously. The probability of a trasmissio by the AP i ay arbitrary slot is give by: P tr1(ap) = (2) This trasmissio is successful oly if oe of the other STAs trasmit i the same slot. Thus, the probability of success give that the trasmissio has occurred is: τ ) P s1(ap) = = ) (3) P -1 τ ) τ (1- ). (6) P s2(ap) = 1. (7) The time for which the chael is busy (we have to cosider the time spet by the STAs as well) is give by: T s2 (AP) = O rts + {H + E p + + SIFS + H + E p + + SIFS + ACK + + DIFS} (8) We ow proceed with the calculatios for the STAs. STAs ca oly trasmit by cotedig ad wiig the chael; ad the probability that at least oe STA trasmits is give as P tr ( STA) Each time a STA udergoes a successful trasmissio, the AP piggybacks aother packet immediately ad hece the total time for successful trasmissio accouts for both the packet duratios ad is the same as T s2(ap) T s(sta) = T s2(ap). (11) + c1(ap) For the AP, the time spet i useful trasmissio cosiderig both the cases i which trasmissios occur is give by: T (E ) = P tr1(ap) P s1( AP) E p + P tr2(ap) P s2(ap) E p τ. (14) Similarly, evaluatig the useful trasmissio time for the remaiig STAs we have, T (E ) = τ. (15) P tr (STA) = [1- ) ] (9) The probability of successful trasmissio give that there is trasmissio from a STA is the probability that the AP did ot trasmit ad oly oe STA trasmitted, -1 ) τ ) τ ) P s (STA) = = (10) 1 ) Similarly, the time spet i a collisio i this case is same as the case 1 whe the AP coteds for the chael. T c(sta) = T c1(ap). (12) The time duratio i which the system is active is give by: + 1 T total = ) σ + τ ) T + s1(ap) τ ) Ts(STA) + 1 [ 1 (1 τ ) ( + 1) τ ) ] T (13) AP p = [ τ ) + τ ) ] Ep = [ ) ( + 1)] Ep P STAs p tr(sta) P s(sta) E p -1 = ) E p We fially calculate the AP s throughput usig (1), (13) ad (14) as: [ τ ) ( + 1)] Ep SAP =. (16) T Total

5 (a) Throughput compariso of AP ad STAs with IEEE DCF (b) Throughput compariso of AP with BDCF (c) Throughput compariso of STAs with BDCF Figure 5. Aalytical ad Simulatio throughput comparisos of BDCF T s = T s1(ap) ad T c = T c1(ap) (18) Let P tr deote the probability of at least oe trasmissio i a cosidered slot time ad P s deote a successful trasmissio. As the AP ad the STAs coted for the chael with equal privileges, the share of the AP i the system throughput is S /( + 1) while that of the remaiig STAs is S [ /( + 1) ] where S is the system throughput as derived by Biachi [14]: tr s p S = (19) From (1), (13) ad (15) the throughput of the STAs is give by: -1 τ ) Ep Sode = (17) TTotal We ext evaluate the throughput for the STA ad AP for DCF based o the model preseted i [21]. The time, T s, for which chael will be busy due to a successful trasmissio is same as T s1(ap). Also the collisio time remais costat i both BDCF ad DCF. Thus, (1- P ) + P P T + (1 P )P T tr P P E tr s s Figure 5(a-c) shows the umerical results of the throughput as a fuctio of umber of statios ad validate our aalysis through simulatios. We cosider 1 Mbps chael ad other parameters are same as described i ext sectio. Figure 5(a) shows the obtaied throughput by AP ad STAs with DCF. Clearly, we ca observe a substatial ufairess betwee the throughputs of AP (dowlik flows) ad STAs (uplik flows). This is because of the CSMA/CA mechaism, which provides equal access privileges to the AP ad STAs. Thus the AP oly gets 1/(+1) of the total available badwidth (with STAs ad a AP), while the STAs obtai a higher share (/(+1)). O the other had, with the preferetial treatmet for AP i BDCF the AP achieves fair share of the badwidth (almost equal to that obtaied by the STAs). 5. SIMULATION RESULTS I this sectio, we compare the performace of BDCF with DCF ad DCF+ [5]. We have implemeted BDCF ad DCF+ i the s-2 simulator (versio 2.26) [6]. For our simulatios, we have used the sceario described i Figure 1, s tr c where a AP is servig 10 statioary STAs. We placed the STAs such that all were i the trasmissio rage of each other, thus avoidig ay hidde termial problem. The badwidth of the wireless chael is set to 2 Mbps ad the AP is coected to a server through a wired lik with badwidth of 100 Mbps ad 2ms of propagatio delay. We have cosidered both symmetric ad asymmetric traffic patters: For symmetric traffic we cosidered 5 uplik ad 5 dowlik flows, deoted by 5Up-5D ad for asymmetric traffic we agai cosider: 3 uplik flows ad 7 dowlik flows, deoted by 3Up-7D. We assume that o two flows start/ed at the same ode. We have doe a comparative aalysis of the aggregate throughput, ad per stream fairess of IEEE DCF, BDCF ad DCF+ for the aforemetioed traffic patters. All results preseted here are averaged over 10 simulatio rus with differet seed values. 5.1 Aalysis of UDP traffic Aggregate Throughput We first compare the performace of the aggregate uplik/dowlik throughputs achieved by the three schemes (BDCF, DCF ad DCF+). We vary the DATA rate of the CBR traffic ruig over UDP from 100Kbps to 800Kbps ad measure the aggregate throughput achieved by the flows i either directios (uplik ad dowlik). Cosiderig the two differet traffic patters metioed above, Figures 6(a-b) ad 6(c-d) show the aggregate uplik ad dowlik throughputs respectively. We observe a huge differece betwee the uplik ad dowlik throughputs whe DCF/DCF+ is employed. As show i Figure 6(a-b), at low loads all the three protocols have similar performace. This is because the chael is relatively free ad all odes ca trasmit their packets without much cotetio. With DCF/DCF+, as we icrease the traffic load, the aggregate uplik throughput rapidly icreases ad dowlik flows starve. Eve for margial icrease i the traffic load, with IEEE DCF or DCF+, the uplik flows obtai a very high throughput ad completely domiate the access to the chael

6 Throughput (Kbps) 900 Uplik 800 Dowlik BDCF DCF DCF+ 6(a) 3Up-7D 6(b) 5Up-5D 7(a) Asymmetric traffic (U3-D7) 800 Uplik 700 Dowlik Throughput (Kbps) BDCF DCF DCF+ 6(c) 3Up-7D 6(d) 5Up-5D 7(b) Symmetric traffic (U5-D5) Figure 7. TCP Aggregate Figure 6. Aggregate Throughput for UDP Traffic Throughput Dowlik flows experiece severe cogestio resultig i drastic reductio of the aggregate dowlik throughput. Oe importat thig to ote from Figure 6(d) is that, till 200Kbps load, the throughput of dowlik traffic icreases as there is sufficiet badwidth to accommodate all uplik ad dowlik flows. O the other had, with BDCF the available badwidth is fairly distributed amog uplik ad dowlik traffic. The dowlik flows achieve a fair share of the total badwidth, without affectig the performace of uplik flows. This ca be attributed to the preferetial treatmet for the AP. AP obtais substatial chael share to accommodate all the dowlik flows ad thus limitig the uplik traffic from the odes. I DCF, the throughput of idividual uplik flows is early 4 times the throughput of idividual dowlik flows. However whe BDCF is employed, this sharp differece does ot occur. The throughput of idividual uplik flows is almost equal to ay of the idividual dowlik flows. The aggregate dowlik throughput is improved by early 300% whe compared with DCF/DCF+. Hece, by providig preferetial access to the AP, BDCF esures fair sharig of badwidth amog uplik ad dowlik flows. We also aalyzed the performace of TCP traffic both the traffic patters. Oce agai with DCF/DCF+ we observe similar ufairess amog the uplik/dowlik flows. Figure 7 shows the throughput performace of TCP traffic. From figures 7(a-b), we ca observe the disgraceful performace of the dowlik TCP traffic flows whe legacy DCF is employed. I cotrast, whe BDCF is employed, we observe fair sharig of the available badwidth amog uplik ad dowlik flows for all traffic patters. For istace, with BDCF i the asymmetric traffic (3Up-7D) sceario, the dowlik flows achieve Kbps (120 Kbps per flow); while the uplik flows obtai aroud 389 Kbps (129 Kbps per flow). However, with DCF ad asymmetric traffic (3Up- 7D), the dowlik flows get oly a meager 610 Kbps (aroud 87 Kbps per flow), while the aggregate uplik throughput (Figure 7(b)) is as high as 470 Kbps (aroud 157 Kbps per flow). The poor performace of TCP with DCF ca be explaied as follows. TCP geerates bi-directioal traffic. TCP ACKs of the uplik flows ad TCP DATA packets of the dowlik flows compete for the limited chael share of the AP. This icreased dowlik traffic load at the AP leads to overflowig of the lik layer queues at the AP ad resultig i excessive packet drops. TCP cogestio cotrol algorithm further worses the situatio by reducig the cogestio widow of a flow whe it detects a packet drop. If a TCP DATA packet of a dowlik flow is dropped at the AP, timeout occurs at the source ad hece the cogestio widow is decreased, leadig to a lower throughput. However, if a TCP ACK of a uplik flow is dropped, due to the cumulative ature of TCP ACKs, evetually aother ACK with a higher sequece umber will reach the uplik flow source ad the TCP cogestio widow will ot be reduced. Thus the dowlik traffic experieces severe

7 8(a) UDP Traffic (3Up-7D) Figure 8. Fairess Idex 8(b) UDP Traffic(5Up-5D) cogestio cotrol while the uplik traffic is ot affected at the same rate. Cosequetly the uplik flows ejoy higher et throughputs whe compared to dowlik flows. I summary, dowlik flows are ot oly affected by the limited chael availability for the AP, but also due to TCP-DATA packet drops at the AP. We carefully observed the cogestio widow growth of all flows ad otice that the cogestio widow of dowlik flows experiece frequet cut dow due to packet drops at the AP (ad thereby leadig to timeouts). I cotrast, the cogestio widow for all the uplik flows grows cotiuously. Similar observatio is also reported by the authors i [4]. Oce agai BDCF overcomes this problem by givig a preferetial treatmet to the AP. Wheever a packet from the STAs is received by the AP; it ca immediately trasmit a packet (recall from sectio 3) from its buffer without ay eed for chael cotetio. I this way, at high traffic loads, the AP ca avoid large queues ad overflows by immediately trasmittig either the TCP- ACKs/TCP-DATA packets. As more ad more dowstream DATA packets are trasmitted the dowlik flows achieve acceptable throughputs without affectig the uplik flows Fairess Idex I this sectio we compare the fairess amog all the flows. We have used the Jai s fairess idex (f) [8] to measure the fairess amog the flows. It is give by: f = * x i i = 1 i = 1 x 2 2 i 1 i, Where, there are flows i the etwork ad x i is the throughput achieved by flow i. The fairess idex is always positive ad whe it approaches oe, it implies that all the flows are gettig equal share of the available badwidth. Whe the fairess idex drops or has egative slope, the it idicates that the available badwidth is ot fairly shared amog the flows. The fairess idexes for asymmetric (3Up-7D) ad symmetric traffic (5Up-5D) cofiguratios are show i Figure 8(a) ad 8(b), respectively. The fairess idex is ear costat ad is close to oe with BDCF i both traffic cofiguratios. This is because all the flows obtai a fair share of the available badwidth. However, this is ot the case whe legacy DCF ad DCF+ are employed, as the dowstream flows achieve very low throughputs whe compared to the throughputs of uplik flows. 6. RELATED WORK Fairess provisioig i wireless etworks has bee a attractive area of research ad has bee explored at various layers. Most of the research addresses the problem of ufairess observed i the upper layers. The ufairess problem betwee uplik ad dowlik TCP flows i a AP based WLAN was iitially reported by Ramjee et al. [2].They studied the iteractio betwee TCP ad IEEE MAC protocol ad idetified the buffer size at the AP as the cause for ufairess, the they proposed receiver widow size maipulatio i the TCP ACK to gover the access of wireless lik. Bottigliego et al. [3] studied the ufairess due to chael uavailability. They aalyzed the chael coditios betwee the STA ad the AP, choosig the STA with best chael coditio for trasmissio ad compesatig the other STAs later with a burst trasmissio. The cocept of reverse chael reservatio i WLANs has bee first itroduced i the DCF+ scheme [5]. The goal of DCF+ is to improve the performace of TCP i WLANs (usig DCF) by a implicit reservatio of the chael. However, as our simulatio results have idicated, DCF+ also has the ufairess problem because the implicit reservatio is used by ay STA. Recetly, attempts to improve performace of TCP i multi-hop ad hoc etworks by similar techiques were proposed i [10] ad [11]. I [10], the authors try to reduce itra ad iter flow cotetio. Itra flow cotetio is ecoutered betwee TCP-DATA ad TCP-ACK packets of a sigle TCP flow while Iter-flow cotetio is experieced betwee TCP-DATA packets of differet TCP flows. They propose two schemes Quick exchage, i which the seder

8 reserves the reverse chael for the TCP-ACK packets ad Fast-forward, i which the seder allows the receiver to forward the curretly received packet towards the destiatio. Kuag et al. [11] proposed a multi-chael MAC ad reverse chael reservatio for mitigatig the DATA-DATA ad DATA-ACK cotetios. However they have ot addressed the fairess problem for the ifrastructure based etworks. Moreover, it should also be oted that the above schemes are desiged for MANETs. Furthermore, this ability to seak a packet is eabled for ay receivig STA that has a packet to trasmit. Hece, whe applied to AP based WLANs, these schemes ca ot distiguish betwee a AP ad other STAs ad will ot be able provide a preferetial treatmet to AP packets oly. Kim et al [7] propose a utilizatio based uplik/dowlik fairess mechaism. The AP couts the umber of STAs based o the uique MAC addresses ad calculates the utilizatio of uplik ad dowlik traffic. If the dowlik traffic is less i a specific time widow the the AP starts trasmittig the DATA frames after deferrig for PIFS time duratio followig the ACK from a STA. Thus the regular chael access policy of deferrig for DIFS duratio ad exchagig the RTS-CTS is avoided. However i their scheme some dowlik DATA frames may still suffer higher delays as it takes some time to detect the ufairess. I cotrast with our scheme the AP seaks DATA frames wheever a uplik trasmissio is sesed thus achievig higher dowlik throughputs. 7. CONCLUSION We experimetally demostrate the existig ufairess problem i typical WLAN hotpots ad proposed a simple ehacemet to DCF for overcomig the ufairess problem. We developed aalytical models to evaluate the throughputs of AP ad STAs ad verify these models through extesive simulatio study. Our proposed BDCF protocol eables the AP to access the chael more frequetly by gratig a preferetial treatmet. I additio to this, our protocol also reduces the time wasted i chael cotetio ad backoff mechaism at the MAC layer. We otice that the aalytical results fid i good agreemet with the simulatio results. Comprehesive simulatios are also coducted o various traffic patters for both TCP ad UDP traffic. It has bee demostrated that BDCF successfully solves the ufairess problem alog with substatial improvemets i the throughputs ad reductio i the ed-to-ed delays of the dowward traffic (aroud 300%). REFERENCES [1] IEEE Std IEEE Stadard for Wireless LAN Medium Access Cotrol (MAC) ad Physical Layer (PHY) Specificatio, ISO/IEC :1999 (E), Aug., [2] S. Pilosof, R. Ramjee, D. Raz, Y. Shavitt, ad P. Siha, Uderstadig TCP Fairess over Wireless LAN, IEEE INFOCOM, [3] M. Bottigliego, C. Casetti, C. F. Chiasserii ad M. Meo, Short-term Fairess for TCP Flows i b WLANs, IEEE INFOCOM [4] I. Aad ad C. Casette, Differetiatig mechaisms i IEEE , IEEE INFOCOM [5] H. Wu, Y. Peg, K. Log, S. Cheg, ad J. Ma, Performace of Reliable Trasport Protocol over IEEE Wireless LAN: Aalysis ad Ehacemet, IEEE INFOCOM [6] UCB/LBNL/VINT Network Simulator (s-2), Available at [7] S. W. Kim, B.S.Kim, Y. Fag, Dowlik ad uplik resource allocatio i IEEE wireless LANs, IEEE Tras. o Vehicular Techology, Ja [8] R. Jai, The Art of Computer Systems Performace Aalysis, Joh Wiley ad Sos, [9] H. Balakrisha ad V. Padmaabha, How Network Asymmetry affects TCP performace, IEEE commuicatio magazie, 60-67, April [10] D. Berger et al., TCP-Friedly Medium Access Cotrol for Ad-Hoc Wireless Networks: Alleviatig Self-Cotetio, MASS 04, Oct 24-27, [11] T. Kuag ad C. Williamso, A bidirectioal multichael MAC protocol for improvig TCP performace o multihop wireless ad hoc etworks, Proc. of ACM/IEEE MSWiM, October [12] J. Malie, Host AP Driver package for Liux, Available at [13] NLANR/DAST Iperf, Available at [14] G. Biachi, Performace aalysis of the IEEE distributed coordiatio fuctio, IEEE J. Select. Areas Commuicatio, Mar