Modeling e for data traffic parameter design

Transcription

1 Modelng 8211e for data traffc parameter desgn Peter Clfford Ken Duffy John Foy Douglas J Leth and Davd Malone Hamlton Insttute NUI Maynooth Ireland TCP throughput (Mbps Number of the upstream connecton Fg 1 Competng TCP uploads 1 statons (NS2 smulaton 8211 MAC 3s duraton parameters as n Table I Abstract Ths paper ntroduces a fnte load mult-class 8211e EDCF model that s smple enough to be explctly solvable The model s nevertheless flexble enough to model the mpact of 8211e parameters on the prortzaton of realstc traffc We emphasze that a modelng framework whch allows nonsaturated sources s essental n the study of realstc traffc We apply the model to a stuaton of practcal nterest: competng TCP flows n an nfrastructure network The model allows us to make a prncpled selecton of 8211e parameters to resolve problems hghlghted n ths scenaro Model predctons and parameter selectons are valdated aganst smulaton and experment The model s shown to be accurate and the parameters effectve I INTRODUCTION 8211a/b/g has been extremely successful but s not wthout shortcomngs whch has motvated the defnton of the 8211e extensons to the basc 8211 MAC For nstance t s known that cross-layer nteractons between the 8211 MAC and the flow/congeston control mechansms employed by TCP can lead to gross unfarness between competng flows and ndeed sustaned flow lockout (eg [1][2] For example consder an 8211b 11Mbps network consstng of laptops tryng to upload large data fles usng TCP Fgure 1 plots smulated throughput acheved by each staton The exstence of gross unfarness s clearly evdent It s wdely recognzed that the 8211a/b/g MAC requres greater flexblty to allevate dffcultes such as those n the example above and consequently the new 8211e standard allows tunng of MAC parameters that have prevously been constant Although the 8211e standard provdes adustable parameters wthn the MAC layer the challenge s to understand how best to use ths flexblty to acheve enhanced network performance The 8211e MAC has been the subect of emprcal studes (eg [3][4][5] and mult-class 8211e models do already exst (eg [6][7][8] However these models are strctly confned to saturated condtons; that s where every staton always has a packet to send To understand the operaton of 8211e n the context of realstc traffc we argue that saturated models are nadequate and that t s essental to model traffc sources wth fnte (nonsaturated demands For example saturated models are not able to capture the behavor exhbted n the example above Data traffc such as web and emal whch consttutes the vast maorty of traffc on current networks s typcally bursty n nature Even long-lved data traffc such as large fle transfers are problematc for saturated modelng as delayed ackng s ubqutous n TCP recevers and means that TCP ACK (acknowledgement transmsson s nonsaturated even f the TCP sender s tself saturated In the context of traffc prortzaton we note that when hgh prorty traffc s lowrate or on-off ths leads to dfferent prortzaton schemes from stuatons where the hgh prorty traffc s greedy or saturated To see ths observe that when hgh-prorty traffc s saturated strct prortzaton schemes cause hgh-prorty traffc to swamp the network Strct prorzaton (plus admsson control s n contrast a standard approach when hgh prorty traffc s low rate such as voce Note that the saturaton of a wreless staton s logcally dstnct from whether the network s heavly or lghtly loaded It s possble for a network of saturated statons to be lghtly loaded f there are only a small number of statons and conversely a network of nonsaturaton statons may be heavly loaded f there are many statons We also note that nterestng features of 8211 MAC behavor only emerge n nonsaturated condtons For example a saturated model cannot predct maxmum network throughput as t s well known that for CSMA/CA random access schemes of the type used n 8211 the throughput s generally not a monotonc functon of offered load ([9] That s there exsts a pre-saturaton throughput peak Ths occurs n 8211a/b/g [1] and wll be shown n ths artcle to persst n 8211e The man contrbuton of ths paper s a mult-class 8211e EDCF fnte-load model that s smple enough to be explctly solvable but complex enough to accurately predct the throughputs of unsaturated traffc and the mpact of the three most sgnfcant 8211e MAC parameters on traffc prortzaton: TXOP AIFS and CW mn In partcular modelng the effect of AIFS ntroduces dffcultes but ts ncluson s fundamental n understandng the full power of 8211e prortzaton We demonstrate the value of the model by usng t to determne settngs of the 8211e parameters that restore

2 farness to the TCP flows n Fgure 1 II RELATED WORK A number of fnte-load models of the 8211a/b/g DCF exst ncludng [1][11][12][13] [14][15][16][17][18] None of these models support multple traffc classes dfferentated by all the varable 8211e MAC parameters: AIFS CW mn and TXOP As noted prevously [6][7][8] develop models of the 8211e EDCF but these are confned to saturated traffc condtons and thus are unsuted for the desgn of prortzaton schemes under realstc traffc condtons Wth regard to TCP unfarness early work [19] studes the mpact of path asymmetres n both wred and wreless networks More recently [2][21] specfcally consder TCP unfarness ssues n 8211 WLANs All of these authors seek to work wthn the constrants of the basc 8211 MAC not utlzng the flexblty of 8211e In [1][2][22] the authors use 8211e functonalty to restore TCP farness As that work was conducted wthout a fnte load 8211e model the proposed parameter settngs were derved emprcally III IEEE 8211 AND 8211E CSMA/CA The 8211 MAC layer CSMA/CA mechansm employs a bnary exponental back-off algorthm to regulate access to the shared wreless channel On detectng the wreless medum to be dle for a perod DIF S each staton ntalzes a counter to a random number selected unformly n the nterval [ CW 1] Tme s slotted and ths counter s decremented once durng each slot that the medum s observed dle A sgnfcant feature s that the countdown halts when the medum becomes busy and resumes after the medum s dle agan for a perod DIF S Once the counter reaches zero the staton attempts transmsson and can transmt for a duraton up to a maxmum tme TXOP (defned to be one packet wthout 8211e If two or more statons attempt to transmt smultaneously a collson occurs Colldng statons double ther CW (up to a maxmum value select a new back-off counter unformly and the process repeats After successful transmsson CW s reset to ts mnmal value CW mn and a new countdown starts regardless of the presence of a packet at the MAC If a packet arrves at the MAC after the countdown s completed the staton senses the medum If the medum s dle the staton attempts transmsson mmedately; f t s busy another back-off counter s chosen from the mnmum nterval Ths bandwdth savng feature s called post-back-off The new 8211e MAC enables the values of DIF S (called AIFS n 8211e CW mn and TXOP to be set on a perclass bass for each staton That s traffc s drected to up to four dfferent queues at each staton wth each queue assgned dfferent MAC parameter values IV 8211E EDCF FINITE-LOAD MODEL As t wll suffce for the applcatons presented n ths paper we assume there are two AIFS values AIFS 1 and AIFS 2 We dvde our statons nto two classes by AIFS value Wthn each class statons can have dstnct arrval rates CW mn values and so forth Wthout loss of generalty those n the class 1 are assumed to have an AIFS smaller than or equal to those n class 2 Statons n each class are modeled by Markov chans of dstnct structure whose transton probabltes are functons of ther system parameters The Markov chans are coupled by the operaton of the network States n the Markov chan model for class 1 statons are labeled by a par of ntegers ( k or ( k e The varable represents the back-off stage whch s ncremented (to a possble maxmum m when attempted transmsson results n collson and set to when transmsson s successful After attempted transmsson the varable k s chosen randomly wth a unform dstrbuton on the ntegers n the range [ W 1] where W = 2 and s the mnmum contenton wndow Whle the medum s dle k s decremented If a packet s present transmsson s attempted when k = The empty states ( k e represent the staton when t does not have a packet to send After successful transmsson f a hgher layer does not provde a packet the MAC layer contnues to decrement k to If a packet arrves durng the countdown the staton swtches to the approprate ( k state Otherwse f countdown has ended wth no packet the staton s n the state ( e When a hgher layer provdes a packet the staton senses the medum If the medum s sensed dle transmsson s attempted mmedately If the medum s sensed busy a stage back-off s ntated now wth a packet The chan for class 2 statons has to be augmented because ther larger AIFS value results n class 1 statons countng down before class 2 statons treat the medum as dle Let D be the nteger number of slots dfference n the AIFS of class 2 and AIFS of class 1 We model the behavor of a class 2 statons wth a three dmensonal Markov chan ndexed ( k d and ( k d e f the MAC layer s empty e there s no packet n the MAC The varable d { D} represents hold states for class 2 That s d > represents states n whch the class 2 statons cannot decrement k whle class 1 flows do as they are not treatng the medum as dle When n a hold state class 2 statons must count up to D before returnng to a non-hold state wth d = Our man assumptons are the same as n [23][6][1] We assume there are no hdden statons and errors are only caused by collsons Condtoned on attempted transmsson each staton has a fxed probablty of collson rrespectve of the network s hstory In addton as n [1][11][12] for each staton there s a fxed probablty of a packet arrvng to the MAC durng transtons n the Markov chans In Secton IV-C we relate the model arrval probablty to the real offered load In the followng two subsectons we defne the transton probabltes for each staton s Markov chan Wthn a gven class these chans have the same structure Calculatons based on ther statonary dstrbutons lead to the equatons n Secton IV-C that determne the model s soluton Let n 1 be the number of statons n class 1 and n 2 the number n class 2 We denote by p (1 {1 n 1 } the probablty that staton n class 1 wll experence a collson gven t s attemptng transmsson and by q (1 be

3 the probablty the MAC receves a packet durng a statetransmsson n the chan We defne p (2 and q (2 {1 n 2 } smlarly for class 2 staton We denote the probablty that staton n class 1 attempts transmsson by τ (1 and by τ (2 the probablty that staton n class 2 attempts transmsson condtoned on t not beng n hold state For notatonal convenence we suppress subscrpts when descrbng each ndvdual staton s Markov chan A Class 1 statons Markov chan For a staton n class 1 let p be the probablty of collson gven attempted transmsson τ be the probablty of transmsson and q be the probablty a hgher layer presents a packet to the MAC The transton probabltes of a class 1 staton s Markov chan are lsted n full below They are determned by straght-forward logc For < k < W < m we have P (( k 1 ( k = 1 P (( k 1 e ( k e = 1 q and P (( k 1 ( k e = q For m and k we have P (( k e ( = ((1 p(1 q/ P (( k ( = ((1 pq/ and P ((mn( + 1 m k ( = p/w mn(+1m The most complex transtons occur from the ( e state where P (( e ( e = 1 q + q(1 p(1 p < k < P (( k e ( e = q(1 p(1 p k < W 1 P ((1 k ( e = q(1 pp W 1 k < P (( k ( e = qp B Class 2 statons Markov chan We begn by dentfyng the probablty that ths class 2 staton observes the medum s slent wth the probablty that t would not have a collson f t attempted transmsson as 1 p = n 1 (1 (1 τ (2 (1 τ where the second product s over all class 2 statons other than the one under consderaton Defne P S1 to be the probablty that all class 1 statons are slent n 1 P S1 = (1 τ (1 (1 The transton probabltes of a class 2 staton s Markov chan are lsted n full below We start wth transtons from non-hold states For < k W 1 and > we have P (( k 1 ( k = 1 p P (( k 1 ( k = p P (( k 1 e ( k e = (1 p(1 q P (( k 1 ( k e = (1 pq P (( k 1 e ( k e = p(1 q For k and P (( k 1 ( k e = pq P (( k 1 e ( = P (( k 1 ( = P ((mn( + 1 m k 1 ( = (1 p(1 q (1 pq p W mn(+1m The fnal set of non-hold states we need to consder f the wndow counter reaches and there s stll no packet to send We deal wth them n a way that enables us to gve the explct expresson n Equaton (4 below for the probablty of not beng n a hold state We refne ( k e further nto the states ( k esense and ( k etrans In ( esense the staton has no packet and s sensng f the medum s busy If t s busy t goes to a hold state If t s dle and no packet arrves t remans n ( esense but f a packet arrves t goes to the second new state ( etrans In ( etrans the source transmts Regardless of what happens (collson successful transmsson the state that follows s a hold state The hold states ( k esense and ( k etrans k > are kept separate because f an arrval occurs whle n ( k esense any k a new back-off s ntated on departng from the hold states Ths necesstates the ntroducton of a new arrval probablty q h the probablty a packet arrves at the MAC at some stage durng transtons from ( 1 esense to successful departure from ( D esense It s not necessary to gve an expresson for q h n terms of q as t cancels out before our fnal equatons but smplfes the dervaton Thus P (( 1 esense ( esense = p P (( esense ( esense = (1 p(1 q P (( etrans ( esense = (1 pq k > P (( k 1 e ( etrans = (1 p(1 q P (( k esense ( etrans = P (( k 1 ( etrans = P ((1 k 1 ( etrans = p W 1 (1 p(1 q (1 pq Turnng our attenton to transtons from hold states For k P (( k ( D esense = P S1 1 q h P (( esense ( D esense = P S1 (1 q h P ((1 k ( D etrans = P S1 p W 1 P (( k ( D etrans = P S1 1 p q 1 p k > P (( k e ( D etrans = P S1 (1 q 1 p P (( esense ( D etrans = P S1 (1 q For 1 < D P (( + 1 esense ( esense = P S1 P (( + 1 etrans ( etrans = P S1 For 1 D P (( 1 esense ( esense = (1 P S1 P (( 1 etrans ( etrans = (1 P S1

4 For 1 < D For k > P (( k + 1 e ( k e = P S1 (1 q P (( k + 1 ( k e = P S1 q P (( k 1 e ( k D e = P S1 (1 q P (( k 1 ( k D e = P S1 q For k > 1 D P (( k 1 e ( k e = (1 P S1 (1 q P (( k 1 ( k e = (1 P S1 q For 1 < D k P (( k + 1 ( k = P S1 k > P (( k 1 ( k D = P S1 P ((mn( + 1 m k (1 D = P S1 p W mn(+1m P (( k ( D = P S 1 (1 pq P (( k e ( D = P S 1 (1 p(1 q P (( k 1 ( k = (1 P S1 C Relatng staton and network models Solvng the Markov Chans for ther statonary dstrbutons as outlned n the Appendx leads to a relaton between p and τ Irrespectve of class and staton each trple (p q τ = (p ( q( τ = 1 η τ ( {1 2} and {1 n } s related by ( q 2 (1 q(1 p(1 (1 q W q2 (1 p 1 q where the normalzaton constant η s η = q 1 (1 q + qw(qw+3q 2 + (1 q 2(1 q(1 (1 q + q(w+1(p(1 q q(1 p2 2(1 q ( pq + 2 ((1 q2(1 p 2W(1 p p(2p M 1 1 (1 q (1 W p2 (1 2p + 1 and s the mnmum contenton wndow The hold probablty P h s the same for all class 2 statons because f one them s n a hold state they all are A combnaton of the statonary dstrbuton of class 2 statons Markov chans and the fxed probablty of collson assumpton gves an expresson for P h If D s zero then the hold probablty P h s zero otherwse t s P h = (1 n 1 (1 (1 τ n 2 (2 (1 τ 1 + (1 n 1 (1 (1 τ n 2 (2 (1 τ D P S 1 D P S 1 (2 (3 (4 where P S1 s defned n Equaton (1 From the network model t s possble to deduce the followng non-lnear equatons (5 and (6 that couple all statons n the network Ther soluton completely determnes p ( and τ ( from whch throughputs and other performance metrcs can be determned: for {1 n 1 } p (1 = 1 n 2 (1 τ (1 (P h + (1 P h (1 τ (2 (5 and for {1 n 2 } p (2 n 1 = 1 (1 τ (1 Havng solved for all p ( τ ( throughputs D Throughput (1 τ (2 (6 and P h we can determne staton The length of each state n the Markov chan s not a fxed perod of real tme Each state may be occuped by a successful transmsson a collson or the medum beng dle To convert between states and real tme we must calculate the expected tme spent per state whch s gven by n 1 E s = (1 P tr σ + P (1 s: T (1 where: n 1 + r=2 1 k (1 1 < <k(1 r n n 2 s: + (1 P c:k (1 n 1 (2 P c:k (2 n 2 1 k(1 r P (2 s: T (2 s: T c:k (1 1 k(1 r T c: k (2 s k(2 s s=2 1 k (2 1 k(2 s 1 < <k(2 s n 1 n 2 (1 (2 P r=1 s=1 1 k (1 c: k(1 1 k(1 r Tc: k(1 1 k(1 r 1 < <k(1 r n1 k (2 1 k (2 1 k(2 k (2 s 1 k(2 s 1 < <k(2 s n2 n 1 P tr = 1 (1 τ (1 (1 τ (2 n 2 s the probablty at least one staton attempts transmsson; σ s the slot-tme; n 2 P (1 s: = τ (1 (1 τ (1 (P h + (1 P h (1 τ (2 s the probablty staton n class 1 successfully transmts; T (1 s: s the tme taken for a successful transmsson from staton n class 1; n 1 n 2 P (2 s: = (1 P h τ (2 (1 τ (1 (1 τ (2 = s the probablty staton n class 2 successfully transmts; s the tme taken for a successful transmsson from T (2 s: (7

5 staton n class 2; P (1 c:k (1 1 k(1 r = r τ (1 k (1 k (1 1 k(1 r n 2 (1 τ (1 (P h + (1 P h (1 τ (2 s the probablty that only the class 1 statons labeled k (1 1 to k r (1 experence a collson by attemptng transmsson whle class 2 statons are n a hold state or are not attemptng transmsson; s = (1 P h (1 τ (2 P (2 c:k (2 1 k(2 s n 1 (1 τ (1 τ (2 k (2 k (2 1 k(2 s s the probablty that only the class 2 statons labeled k (2 1 to k s (2 experence a collson by attemptng transmsson whle class 1 are not attemptng transmsson; r s (1 (2 P = (1 P h c: k(1 1 k(1 r k (2 1 k(2 s k (1 1 k(1 r τ (1 k (1 (1 τ (1 τ (2 k (2 k (2 1 k(2 s (1 τ (2 s the probablty that only the class 1 statons labeled k (1 1 to k r (1 and class 2 statons labeled k (2 1 to k s (2 experence a collson by attemptng transmsson; and Tc: k(1 1 k(1 r k (2 1 k(2 s s the tme taken for ther collson For example usng the basc 8211b MAC values found n Table I wth staton havng payload E Bytes and Header = PLCP+MAC+CRC+IP T s: = E +Header+δ+SIFS+ACK+PLCP+δ + AIFS 1 Tc: k(1 1 k(1 r k (2 1 k(2 s = max = k(1 1 k(1 r k (2 1 k(2 s T c: T c: = E +δ+header+sifs+acktmeout where for 8211b ACKTmeout s the tme taken for an ACK plus PLCP plus δ plus DIFS makng T s: = T c: E Relatng offered load to model parameters If a staton s saturated and always has a packet to send then q s 1 If the staton s not saturated then to a frst level of approxmaton q s the probablty that at least one packet arrves at the MAC durng E s s one mnus the probablty that the frst nter-packet tme s greater than E s For example when nter-packet arrval tmes to the MAC are exponentally Duraton(µs Slot tme σ 2 Propagaton delay δ 1 DIFS (AIFS= 5 SIFS (Short Inter Frame Space 1 PLCP 192 MAC Header CRC Header 4 29 IP Header MAC ACK E payload E payload E payload TABLE I 8211B MAC VALUES BASIC RATE 1MBPS AND DATA RATE 11MBPS dstrbuted wth rate λ then q and λ are related by q = 1 exp( λe s Wth packet-length E Bytes the staton s offered load s log(1 q8e 8Eλ = Mbps (8 E s Smlar calculatons are possble wth other nter-arrval tme dstrbutons V MODEL VALIDATION We start wth a representatve selecton of fgures that demonstrate the model s throughput predcton accuracy through comparson wth TU-Berln s [24] 8211e NS2 packet-level smulator Consder a peer-to-peer 11Mbps network whose parameter values are gven n Table I The network conssts of 1 class 1 staton and 2 class 2 statons transmttng 56 Byte packets Each staton n class 2 staton offers 4 tmes the load of each class 1 staton In smulaton nterface buffers are short and packets arrve wth exponentally dstrbuted nter-arrval tmes The model s solved ndependently wth q 1 and q 2 determned by Equaton (8 The graphs n Fgure 2 show throughput for a staton n each class versus offered load The AIFS for class 1 s the 8211b DIF S value; the AIFS for class 2 s DIF S plus D slots of length δ Statons n each class have CW mn = 32 Not only s the accuracy of the model apparent but also the strong prortzaton effect of AIFS The related graphs n Fgure 3 show throughput for a staton n each class versus offered load for a range of mnmum contenton wndow pars For these graphs the AIFS value of each class are the same Prortzaton s stll apparent and agan the model makes remarkably accurate predctons In all presented graphs throughput s not a monotonc functon of offered load that there s pre-saturaton throughput peak and the model predcts ths effect We have performed a large range of valdaton experments matchng smulaton to model predctons and smlar accuracy was obtaned n all of them VI 8211E PARAMETER DESIGN The 8211e extensons to the 8211 DCF allow a number of CSMA/CA parameters to be adusted on a per class bass Our motvaton n developng an 8211e fnte-load model s

6 25 smulaton 25 smulaton 25 smulaton Per staton throughput (Mbps Per staton throughput (Mbps Per staton throughput (Mbps Total offered load (Mbps Total offered load (Mbps Total offered load (Mbps (a D = (b D = 2 (c D = 4 Fg 2 Throughput for a staton n each class vs offered load 1 class 1 statons offerng one quarter the load of 2 class 2 statons Range of D values the dfference n AIFS between class 2 and class 1 (NS2 smulaton and model predctons 8211e MAC 11Mbps PHY 1s duraton MAC parameters as n Table I 25 smulaton 25 smulaton 25 smulaton Per staton throughput (Mbps Per staton throughput (Mbps Per staton throughput (Mbps Total offered load (Mbps Total offered load (Mbps Total offered load (Mbps (a W (1 = 32 W (2 = 16 (b W (1 = 32 W (2 = 64 (c W (1 = 32W (2 = 256 Fg 3 Throughput for a staton n each class vs offered load There are 1 class 1 statons each offerng one quarter the load of 2 class 2 statons Range of CW mn values (NS2 smulaton and model predctons 8211e MAC 11Mbps PHY 1s duraton MAC parameters as n Table I to establsh an analytc bass for the desgn of strateges for adustng the 8211e parameters Before consderng parameter desgn for a stuaton of practcal nterest we frst brefly dscuss the mpact of the mpact of the followng key 8211e parameters: TXOP AIFS and CW mn The effect of TXOP s easy to understand TXOP controls the number of packets (more precsely the tme allowed for packet transmsson that can be sent at a transmsson opportunty Increasng TXOP proportonately ncreases the relatve throughput of statons provdng that they have data to send The effect on performance of the AIFS parameter s more complex To understand the nfluence of the AIFS parameter recall from Secton III that the MAC countdown halts when the wreless medum becomes busy and resumes after the medum s dle agan for a perod AIFS In addton to the ntal delay of AIFS before countdown starts a staton accumulates an addtonal AIFS delay for every packet sent on the medum by other statons leadng to a reducton n the number of transmsson opportuntes that can be ganed by a staton as ts AIFS s ncreased Ths effect s load dependent When the network s lghtly loaded we AIFS dfferences have lttle mpact on throughput However as the network load ncreases statons wth longer AIFS are rapdly penalzed Ths load dependent behavor can be seen by comparng Fgure 2 (a and (b when the network load s low the graphs are qute smlar When the load s hgh there s a dramatc change n the throughput acheved by the lower-rate class The effect s even more apparent n Fgure 2 (c The mpact on throughput of the CW mn parameter s smple When the network s saturated we expect the throughput of a staton to be roughly nversely proportonal to ts CW mn value as CW mn s proportonal to how long a staton must wat between transmssons In Fgure 3 (a (b and (c the CW mn parameters are n the rato 5 2 and 8 respectvely Lookng at the throughput ratos for heave loads we fnd that they are approxmately 4 17 and 7 respectvely confrmng ths ntuton We note that the tunng of the CW mn parameter n the 8211e standard s coarse as the parameter s constraned to be a power of two whch lmts ts utlty One mght expect changng CW mn for 2 statons would have an mpact even n lghtly loaded stuatons due to a combnaton of mpact on collson rates and extra tme spent n backoff However we see from the fgures that there s only a small mpact on throughput untl the network load becomes sgnfcant It s nterestng to note that unlke AIFS 8211e permts the values of CW mn to be reduced below ts default value n 8211 Ths can be used to ncrease the peak throughput of class 2 n Fgure 3 (a beyond that n Fgure 2

7 (a We turn now to the desgn of 8211e parameter selecton strateges for the example from the Introducton preventng lockout of competng flows n data networks A Mtgatng TCP/MAC cross-layer nteractons Exstng work on 8211e s largely drven by qualty of servce requrements of applcatons such as VoIP However network traffc s currently domnated by data traffc (web emal meda downloads etc carred by the TCP relable transport protocol Although lackng the tme crtcal aspect of voce traffc data traffc server-clent applcatons do place sgnfcant qualty of servce demands on the wreless channel Wthn the context of nfrastructure WLANs n offce and commercal envronments there s a requrement for effcent and reasonably far sharng of wreless capacty between competng data flows As noted n the Introducton cross-layer nteractons between the 8211 MAC and the flow/congeston control mechansms employed by TCP leads to gross unfarness between competng flows and even sustaned lockout see Fgure 1 At the transport layer to acheve relable data transfers TCP recevers return ACK packets to the data sender confrmng safe arrval of data packets Durng TCP uploads wreless statons queue data packets to be sent over the channel to ther destnaton and the returnng TCP ACK packets are queued at the wreless AP to be sent to the source staton TCP s operaton mplctly assumes that the forward (data and reverse (ACK paths between a source and destnaton have smlar packet transmsson rates The basc 8211 MAC layer however enforces staton-level far access to the wreless channel That s n statons competng for access to the wreless channel are each able to secure approxmately a 1/n share of the total avalable transmsson opportuntes Wth n wreless statons and one AP each staton (ncludng the AP s able to gan only a 1/(n + 1 share of transmsson opportuntes By allocatng an equal share of packet transmssons to each wreless staton wth TCP uploads the 8211 MAC allows n/(n + 1 of transmssons to be TCP data packets yet only 1/(n + 1 to be TCP ACK packets For larger numbers of statons ths MAC layer acton leads to substantal forward/reverse path asymmetry at the transport layer Asymmetry n the forward and reverse path packet transmsson leads to sgnfcant queueng and droppng of TCP ACKs dsrupts the TCP ACK clockng mechansm hnders congeston wndow growth and nduces repeated tmeouts Repeated tmeouts can lead to a persstent stuaton where flows are completely starved for long perods The requrement s clearly to prortze the access pont (AP so that suffcent bandwdth s avalable for returnng TCP ACKs to be transmtted back to the data statons Followng [1][2][22] we consder prortzng the AP such that TCP ACKs have essentally unrestrcted access to the wreless medum The ratonale for ths approach to dfferentatng the AP makes use of the transport layer behavor Namely allowng TCP ACKs unrestrcted access to the wreless channel does not lead to the channel beng flooded Instead t ensures that the volume of TCP ACKs s regulated by the transport layer rather than the MAC layer In ths way the volume of TCP ACKs wll be matched to the volume of TCP data packets thereby restorng forward/reverse path symmetry at the transport layer When the wreless hop s the bottleneck data packets wll be queued at wreless statons for transmsson and packet drops wll occur there whle TCP ACKs wll pass freely wth mnmal queueng e standard TCP semantcs are recovered Our desgn requrement s that TCP ACK loss rate be no more than around 1 2% We assume that a bandwdth-delay product of bufferng s used at the wreless statons for the data traffc (recall that traffc classes are queued separately n 8211e and t s for example possble for voce traffc to have small buffers and data traffc larger buffers When operatng normally the AIMD acton of TCP congeston control wll mantans a non-empty buffer at each wreless staton whch can therefore be modeled as saturated traffc sources In order to model TCP ACKs arrvng at the access pont we adust the arrval probablty at the access pont to be half the total successful transmsson rate for data packets from the statons Ths models TCP delayed ACKng where a TCP recever sends an ACK packet for approxmately every second data packet Fgure 4 llustrates model predctons the TCP ACK loss rate at the AP and the data throughput acheved by the wreless statons Results are shown for ten wreless statons uploadng data and recevng nonsaturated TCP ACKs back Agan observe that adustment of CW mn alone s unable brng the TCP ACK loss rate below the requred target Use of AIFS alone requres the use of extremely large values (AIFS > 2 slots wth assocated mpact on data throughput (fallng by > 2% We therefore consder the combned use of AIFS and CW mn Values of (CW mn AIFS whch gve a loss rate of close to 1% are (12 (24 (47 (812 Note that data throughput decreases as the value of AIFS ncreases see Fgure 4(a We therefore seek to use a small AIFS value Throughput s however not a monotonc functon of AIFS for small values of AIFS The settngs CW mn = 2 for the AP and AIFS = 4 for the data statons maxmze throughput In the next secton the performance of ths scheme s confrmed by experment VII EXPERIMENTAL SETUP Recently hardware supportng a useful subset of the 8211e functonalty has become avalable Thus we nvestgate the behavor of our parameterzaton n a realstc network rather than through smulatons Our WLAN conssts of a desktop PC actng as an AP and 12 embedded Lnux boxes based on the Soekrs net481 actng as clent statons Each staton s equpped wth an Atheros 8211a/b/g PCI card wth an external antenna The system hardware confguraton s summarzed n Table II Each staton uses a Lnux 2681 kernel and a MADWF wreless drver modfed to allow adustment of the 8211e CW mn AIF S and T XOP parameters Vendor specfc features are dsabled Tests are

8 6 5 W 1 =1 W 1 =2 W 1 =4 W 1 =8 W 1 =16 W 1 = W 1 =1 W 1 =2 W 1 =4 W 1 =8 W 1 =16 W 1 =32 7 Throughput of Data Statons (Mbps Rato of lost traffc to offered load Dfference n AIFS Dfference n AIFS (a Data throughput (b ACK loss Fg 4 1 statons (15 byte packets and AP transmttng (6 byte packets at half acheved data rate (Model predctons 8211e MAC 3s duraton parameters as n Table I Hardware 1 AP Dell GX 28 28Ghz P4 12 staton Soekrs net Mhz 586 WLAN D-Lnk DWL-G52 Atheros AR5212 Buffers default used TCP 64KB 1MB nterface tx 199 packets 1 packets drver tx 2 packets 1 packets TABLE II EXPERIMENTAL SETUP layer congeston control and the standard 8211 MAC n an upload scenaro In ths case the selected parameter sets are demonstrated to be effectve n practce We note that solvng the analytc model s equatons s substantally faster than packet-level smulaton and enables effcent nvestgaton of the 8211e parameter space In ths way the model can be used as a desgn tool to help overcome the standard 8211 MAC s known drawbacks 1 performed usng the 8211b physcal maxmal transmsson rate of 11Mbps wth RTS/CTS dsabled The confguraton of network buffers s detaled n Table II In partcular we have ncreased the sze of the TCP buffers to ensure that we see true AIMD behavor (wth small TCP buffers TCP congeston control s effectvely dsabled as the TCP congeston wndow s determned by the buffer sze rather than the network capacty We have carred out tests nvestgatng the mpact of the sze of nterface and drver queues and obtan smlar results for a range of settngs Further detals of ths setup are descrbed n [22] Expermental results are shown n Fgure 5 wth and wthout prortzaton It can be seen that farness between TCP uploads s restored For other network scenaros the model s predctons can be used n the same way to determne optmal MAC settngs; we have verfed that the suggested values for AIFS and CW mn are a good choce for a broad range of stuatons VIII CONCLUSIONS We have ntroduced an 8211e CSMA/CA model that s smple enough to be explctly solvable but complex enough to accurately predct data throughput We have shown that the model provdes nsght nto the mportance of dfferent 8211e parameters Modelng nonsaturated traffc sources allows us to take an analytc approach to the desgn of prortzaton schemes for practcal stuatons and realstc traffc To demonstrate how the model can be used to make prncpled selecton of 8211e parameters we use the model to resolve serous cross-layer nteractons between transport APPENDIX Frst we make observatons that ad n the determnaton of the statonary dstrbuton to enable us relate p and τ for a class 1 staton Wth b( and b( e denotng the statonary probablty of beng n states ( and ( e as b s a probablty dstrbuton we have m = W 1 = W 1 b( + = b( e = 1 (9 We wll wrte all probabltes n term of b( e and use the normalzaton n Equaton (9 to determne b( e We have the followng relatons To be n the sub-chan (1 the followng must have occurred: a collson from state ( or an arrval to state ( e followed by detecton of an dle medum and then a collson so that b(1 = b( p + b( e q(1 pp For > 1 we have b( = p 1 b(1 and so b(1 b( = 1 p = b( p + b( eq(1 pp (1 1 p 1 The keystone n the calculaton s then the determnaton of b( 1 e Transtons nto ( 1 e from ( e occur f there s an arrval the medum s sensed dle and no collson occurs Transtons nto ( 1 e also occur from ( f no collson and no arrval occurs b( 1 e = b( e q(1 p 2 + (1 p(1 q b( (11 1 Ths work was supported by Scence Foundaton Ireland grant 3/IN3/I396

9 TCP Throughput (Mbps TCP Throughput (Mbps Number of the upstream connecton Number of the upstream connecton Fg 5 Competng TCP uploads 12 statons experment wthout and wth prortzaton (8211e MAC 3s duraton Combnng Equatons (1 and (11 gves b( 1 e = b( e b( = 1 q q b( e (1 pq(1 pq 1 (1 q q (1 p(1 pq(1 (1 q + b( 1 q We then have for 1 > > b( e = (1 qb( + 1 e +b( 1 e wth b( e on the left hand sde replaced by qb( e f = Straght forward recurson leads to expressons for b( e n terms of b( e and b( and we fnd ( (12 Thus the second sum n Equaton (9 1 = b( e = b( e (q /(1 (1 q W The ( chan can then be tackled startng wth the relaton b( 1 = b( (1 pq W 1 = + b( e qp Recurson leads to [ q b( = b( e 1 q ( q 2 1 (1 q p(1 q q(1 p 2 + q(q + q 2 2(1 (1 q + 1 q ] Usng Equaton (12 we can determne b(1 n terms of b( e : pq 2 ( b(1 = b( e 1 q 1 (1 q W (1 p 2 Fnally the normalzaton (3 gves 1/b (e = (1 q + q2 (+1 ( + q(w+1 2(1 q + 2(1 q(1 p 2(1 (1 q q 2 1 (1 q + p(1 q ( q(1 p 2 pq 2 1 (1 q (1 W p2 1 p p(2p (2W m 1 1 2p + 1 (13 The man quantty of nterest s τ the probablty that the staton s attemptng transmsson A staton attempts transmsson f t s n the state ( (for any or f t s n the state ( e a packet arrves and the medum s sensed dle Thus τ = q(1 pb( e + b( whch reduces to Equaton (2 where 1/b( e = η gven n Equaton (13 so that τ s expressed solely n terms of p q and m For class 2 statons τ 2 s the probablty that the staton wll attempt transmsson n a typcal slot condtoned on t not beng a hold state Wth ths condtonng n force the class 2 statons Markov chan s of dentcal structure to that of class 1 Thus the same relatonshp Equaton (2 holds between p and τ for both classes We do however have to consder the statonary dstrbuton of the class 2 chan to calculate P h the probablty that a class 2 staton s n a hold state Let c( k c( k e c( k esense and c( k etrans denote the statonary dstrbuton of the class 2 Markov chan Our ntroducton of the states ( k esense and ( k etrans allows for a smple deducton of P h It enables the dvson n states nto those n whch a source attempts transmsson whch we call frng states {( ; } ( etrans and all other states whch we call non-frng states We establsh relatons that wrte the statonary probablty of hold-states n terms of non-hold states Consder states that have a packet Frstly for those where we are not n a hold state: c(1 = c( p + c( esense q(1 pp; > c( + 1 = c( p; > c( = c( (W /W ; and for > c( = W (c( + (q(1 pq + pq h c( esense + q m=d n m= c( n m e Next we consder states where we have a packet and are n a hold state For non-frng non-stage backoff states wth > and 1 k < D c( k = P S1 c( k 1 and c( 1 = c( p + (1 P S1 D c( k For frng non-stage backoff states wth > and 1 k < D c( k = P S1 c( k 1 and c( 1 = c( + (1 P S1 D c( k For non-frng stage backoff states wth > and 1 k < D c( k + 1 = P S1 c( k + qp S1 c( k e c( 1 = c( p + (1 P S1 D c( k +pqc( e +q(1 P S1 D c( k e Frng stage backoff states wth 1 k < D c( 1 = c( + (1 P S1 D c( k c( k + 1 = P S1 c( k and c( k + 1 e = P S1 c( k e Next consder states that don t have a packet Wth 1 and

10 1 k < D c( 1 e = (1 q c(e(1 pq+c( c( k + 1 e = (1 qp S1 c( k e c( 1 e = (1 qpc( e +(1 q(1 P S1 D c( k e and wth < c( e = n=1 Snce for 1 k < D (1 q n (1 q c( e(1 pq + c( c( k + 1 etrans = P S1 c( k etrans c( 1 etrans = c( etrans +(1 P S1 D c( k etrans c( k + 1 esense = P S1 c( k esense c( 1 esense = pc( esense +(1 P S1 D c( k etrans and for > c( k e + c( k = vol 22 no 5 pp P S1 (c( k 1 + c( k 1 e [8] Z Kong D HK Tsang B Bensaou and D Gao Performance analyss c( 1 e + c( 1 = p(c( e + c( +(1 P S1 D (c( k + c( k e and D (c( k e + c( k = D p(c( + c( e P D+1 k S 1 Therefore we have D c( k = D p (P S1 c( k By smlarly consderatons we have D c( k = D p (P S1 c( for > > and D k c( k = D 1 (P S1 c( for > Thus usng the normalzaton k of the statonary dstrbuton we have an expresson that does not nclude hold states Moreover the frst term conssts of non-hold non-frng states and the second term conssts of nonhold frng states: 1 = (c( esense + c( > + ( D 1 + c( e 1 + p >(c( (P S1 k + (c( etrans + ( D 1 c( 1 + (P S1 k Defnng C non frng to be the probablty of beng n a non-frng non-hold state C non frng := c( esense + > c( + > (c( + c( e C frng to be the probablty of beng n a frng state C frng := c( etrans + c( and the sum S := D (P S 1 k we have 1 = C non frng [1 + ps] + C frng [1 + S] (14 As τ = C frng /(C frng + C non frng and 1 P h = C frng + C non frng Dvdng Equaton (14 by C frng + C non frng and usng the expressons for 1 P h and τ we have P h = 1 (τ(1+s+(1 τ(1+ps 1 Recallng P S1 = (1 τ 1 n1 and 1 p = (1 τ 1 n 1 (1 τ 2 n2 1 leads to Equaton (4 REFERENCES [1] DJ Leth and P Clfford Usng the 8211e EDCF to acheve TCP upload farness over WLAN lnks n WOPT Trento Italy 25 [2] DJ Leth and P Clfford TCP farness n 8211e WLANS n IEEE WrelessCom 25 Mau Hawa USA 25 [3] Qang N L Romdhan and T Turlett A survey of QoS enhancements for IEEE 8211 wreless LAN Wreless Communcatons and Moble Computng vol 5 no 4 pp [4] P Gopalakrshnan D Famolar and Toshkazu Kodama Improvng WLAN voce capacty through dynamc prorty access n IEEE GLOBECOM 24 vol 5 pp [5] Y Xao H L and S Cho Protecton and guarantee for voce and vdeo traffc n IEEE 8211e wreless LANs n IEEE INFOCOM 24 vol 3 pp [6] R Battt and Bo L Supportng servce dfferentaton wth enhancements of the IEEE 8211 MAC protocol: models and analyss Tech Rep DIT-3-24 Unversty of Trento 23 [7] JW Robnson and TS Randhawa Saturaton throughput analyss of IEEE 8211e enhanced dstrbuted coordnaton functon IEEE JSAC of IEEE 8211e contenton-based channel access IEEE JSAC vol 22 no 1 pp [9] D Bertsekas and R Gallager Data Networks Prentce Hall 1987 [1] K Duffy D Malone and DJ Leth Modelng the 8211 Dstrbuted Coordnaton Functon n non-saturated condtons IEEE Communcatons Letters vol 9 no 8 pp [11] D Malone K Duffy and DJ Leth Modelng the 8211 Dstrbuted Coordnaton Functon wth heterogenous fnte load n RAWNET Trento Italy 25 [12] P Clfford K Duffy DJ Leth and D Malone On mprovng voce capacty n 8211 nfrastructure networks n IEEE WrelessCom 25 Mau Hawa USA 25 [13] G-S Ahn A T Campbell A Veres and L-H Sun Supportng servce dfferentaton for real-tme and best-effort traffc n stateless wreless ad hoc networks (SWAN IEEE Transactons on Moble Computng vol 1 no 3 pp [14] M Ergen and P Varaya Throughput analyss and admsson control n IEEE 8211a ACM-Kluwer MONET Specal Issue on WLAN Optmzaton at the MAC and Network Levels 24 [15] Bo L and R Battt Analyss of the IEEE 8211 DCF wth servce dfferentaton support n non-saturaton condtons Lecture notes n Computer Scence vol 3266 pp [16] GR Canten Qang N C Barakat and T Turlett Performance analyss under fnte load and mprovements for multrate 8211 Computer Communcatons vol 28 no 1 pp [17] AN Zak and MT El-Hadd Throughput analyss of IEEE 8211 DCF under fnte load traffc n Frst Internatonal Symposum on Control Communcatons and Sgnal Processng 24 pp [18] L Bonon M Cont and E Gregor Runtme optmzaton of IEEE 8211 wreless LANs performance IEEE Transactons on Parallel and Dstrbuted Systems vol 15 no 1 pp [19] H Balakrshnan and V Padmanabhan How network asymmetry affects TCP IEEE Communcatons Magazne pp [2] A Dett E Grazos V Mnchello S Salsano and V Sangregoro TCP farness ssues n IEEE 8211 based access networks 25 [21] S Plosof R Ramee Y Shavtt and P Snha Understandng TCP farness over wreless LAN n INFOCOM San Francsco USA 23 [22] ACH Ng D Malone and DJ Leth Expermental evaluaton of TCP performance and farness n an 8211e test-bed n ACM SIGCOMM Workshops 25 [23] G Banch Performance analyss of IEEE 8211 dstrbuted coordnaton functon IEEE JSAC vol 18 no 3 pp [24] S Wethölter and C Hoene Desgn and verfcaton of an IEEE 8211e EDCF smulaton model n ns-226 Tech Rep TKN-3-19 Technsche Unverstät Berln 23