An Analytic Study of Tuning Systems Parameters in IEEE e Enhanced Distributed Channel Access

Transcription

1 An Analyic Sudy of Tuning Sysems Parameers in IEEE 80.11e Enhanced Disribued Channel Access Ye Ge Dep of Elecrical Engineering Ohio Sae Universiy Columbus, OH Jennifer C. Hou D of Compuer Science Univ. of Illinois a Urbana-Champaign Urbana, USA jhou@cs.uiuc.edu Sunghyun Choi School of Elecrical Engineering Seoul Naional Universiy Seoul, Korea schoi@snu.ac.kr Absrac In his paper, we derive, based on he analyical model developed by Cali e al., a muli-class model o sudy how o adapively une all he parameers in IEEE 80.11e EDCA and suppor service differeniaion in WLANs. Through analyical modeling, we demonsrae ha by assigning appropriae differen ransmission probabiliies (or conenion window sizes) o saions of differen classes, i is feasible o provide (proporional) service differeniaion and achieve pre-specified argeed hroughpu raios among differen classes, while a he same ime, maximizing he oal sysem capaciy. We also exend he derived heoreical model o analyze he role of AIFS and TXOP values on service differeniaion perceived by differen raffic classes. We show ha, o achieve QoS guaranees (i.e. hroughpu differeniaion) and high channel uilizaion, i may no be desirable o allow uning of muliple parameers (e.g., boh he conenion window sizes and he AIFS values). Insead, he design dimension should be kep small by allowing urning of only one se of parameers, while keeping he oher wo ses of parameers for all he access caegories fixed a he same value (i.e., seing he AIFS values of all access caegories o, which is equivalen o ). We also elaborae on several implemenaion issues on incorporaing heoreical resuls ino IEEE 80.11e. These include (i) how o reduce he compuaional complexiy and pracically calculae resuls on-line, (ii) how o conver he opimal parameers derived in he model ha characerizes he -persisen version of IEEE 80.11e o hose in IEEE 80.11e (which is based on he noion of he conenion window o deermine wheher or no o ransmi in a slo), and (iii) how o on-line measure parameers needed for calculaing he bes value of he conenion window size. Boh he analyical models and he proposed approaches for pracically incorporaing heoreical findings ino IEEE 80.11e EDCA are validaed hrough deailed ns- simulaions and empirical experimenaion on a Linux-based MADWifi driver for wireless LAN devices wih he Aheros chipse.

2 I. INTRODUCTION IEEE based wireless Local Area Neworks (WLANs) have become popular a an unprecedened rae. Besides convenional Inerne applicaions such as , file ransfer, and web access and browsing, WLANs are also expeced o suppor QoS-cenric applicaions such as audio/video sreaming in smar home environmens. The laer applicaions are delay sensiive and require cerain level of hroughpu and delay guaranees. This creaes a urgen need for supporing QoS in based WLANs. The curren IEEE sandard [] defines wo access mehods: (i) he Disribued Coordinaion Funcion (DCF), also known as he basic access mehod, is a carrier sense muliple access proocol wih collision avoidance (CSMA/CA); (ii) he Poin Coordinaion Funcion (PCF) is a polling-based access mehod and uses a poin coordinaor o arbirae access among saions. In DCF, all he daa raffic is ransmied on a firs come firs serve, bes-effor basis. There is no noion of prioriies and all he saions in he basic service se (BSS) conend for he wireless medium wih he same prioriy. When he number of saions in a BSS increases, he probabiliy of collisions increases, leading o frequen reransmission and a decrease in he overall hroughpu (and QoS) [8]. PCF, on he oher hand, was designed o suppor ime-bounded raffic, and defines wo periods beween ransmission of wo consecuive Delivery Traffic Indicaion Message (DTIM) beacon frames: Conenion Free Period (CFP) and Conenion Period (CP). Beacon frames are sen periodically by he access poin (AP), and carry synchronizaion and nework (BSS) informaion. In paricular, DTIM beacon frames are used o indicae he sar of a CFP. During a CP all he saions conend for he wireless medium using DCF, while during a CFP he AP schedules ransmissions o and/or from individual saions using a polling mechanism. In spie of he inenion o suppor ime-bounded raffic, many inadequacies have been idenified in he design [15]: (i) unpredicable beacon delays resuling in significanly shorened conenion free periods (CFPs), and (ii) unknown ransmission duraions of polled saions making i very difficul for he AP o predic and conrol he polling schedule for he remainder of he CFP. In addiion, here has no managemen inerface defined o se up and conrol PCF operaions and o communicae QoS requiremens o he AP. As a resul, i is difficul, if no impossible, o se up a PCF policy which is compaible o raffic policies in he Inerne InServ/DiffServ archiecures. To deal wih he inadequacy of legacy IEEE DCF and PCF access mechanisms in supporing QoS, several schemes have been proposed in lieraure ha aim o enhance IEEE DCF and provide cerain service differeniaion beween differen raffic classes. Schemes proposed in [1, 13, 7] une he values of parameers in DCF (,, iner-space values, and maximum frame lenghs) o achieve service differeniaion. Schemes proposed in [6, 7, 3, 30, 31] adapively adjus he ransmission probabiliy (or he conenion window size) o boh differeniae services among differen raffic classes and o improve channel uilizaion. Vaidya e al. [6] and Banchs e al. [] apply he concep of fair queueing o achieve differeniaed channel bandwidh disribuion among differen raffic classes. Qiao e al. [], Li and Baii [17], and Xiao [9] exend he Markov chain model in [5,3,] o he case of muliple prioriy classes. Based on he exended 1

3 analyical model, Qiao e al. hen proposes a prioriy based fair medium access conrol proocol P-MAC, Li and Baii sudy how o scale he minimum conenion window size and he lengh of he packe payload according o he prioriy of each raffic flow under he sauraed scenario, while Xiao [9] incorporaes hree adjusable parameers (he iniial conenion window size, he rery limi, and he backoff window-increasing facor) ino he model. On he indusry side, IEEE has charered a Working Group E o invesigae possible enhancemen o IEEE MAC. In paricular, he IEEE 80.11e draf sandard defines a new coordinaion funcion called he Hybrid Coordinaion Funcion (HCF). HCF includes boh a conenion-based channel access mehod, called he Enhanced Disribued Channel Access (EDCA) mechanism, for conenion based daa ransmission, and a conrolled channel access, referred o as he HCF Conrolled Channel Access (HCCA) mechanism, for conenion free daa ransmission. A a very high level, EDCA envisions an archiecure of muliple access caegories and assigns o differen access caegories differen values of he minimum idle delay before conenion (AIFS), he minimum and maximum conenion windows ( and ), and he ransmission opporuniy limi (TXOP). Is performance wih respec o service differeniaion has been sudied via simulaion in [1, 0, 18, 19, 1]. (We will give a more deailed summary in Secion II.) Mos of he aforemenioned research/indusry effors on providing QoS in legacy IEEE DCF [1, 13, 7, 3, 31, 6, 7, 30, 6] and on evaluaing IEEE 80.11e [1, 0, 18, 19, 1] are based on simulaion. The only excepion is perhaps hose repored in [, 9, 17]. In paricular, he analyical model derived in [9] can be used o sudy he performance of IEEE 80.11e in erms of he sauraion hroughpu, he sauraion delay and he frame dropping probabiliy. However, he imporan unable parameer in he IEEE 80.11e draf sandard, AIFS, is no figured ino he model. (As a maer of fac, a deailed analysis on how AIFS values affec service differeniaion among differen raffic classes has been lacking.) The issue of on-line fine uning he above parameers o mee he QoS requiremens of differen access caegories in he presence of nework dynamics is also no discussed in [, 9]. In his paper, we bridge he gap and derive, based on he analyical model developed in [8], a muli-class model o sudy how o adapively une parameers in IEEE 80.11e EDCA and suppor service differeniaion in WLANs. We only consider EDCA, because EDCA bears some similariies o DCF and here are several analyical models ha characerize daa ransmission aciviies governed by DCF and for which we will leverage. (We leave he ask of analyzing daa ransmission aciviies governed by HCCA and uning heir parameers as par of our fuure work.) Through analyical modeling, we demonsrae ha by assigning appropriae differen ransmission probabiliies (or differen conenion window sizes in he case of conenion window-based IEEE 80.11e) o saions of differen classes, i is feasible o provide service differeniaion and achieve pre-specified argeed hroughpu raios among differen classes, while a he same ime maximizing he oal sysem capaciy. We also exend he derived heoreical model o incorporae he effec of using differen AIFS and TXOP values on service differeniaion perceived by differen raffic classes. We show

4 ha, o achieve QoS guaranees (i.e. hroughpu differeniaion) and high channel uilizaion, i may no be desirable o allow uning of muliple parameers (e.g., boh he conenion window sizes and he AIFS values). Insead, he design dimension should be decreased by allowing urning of only one se of parameers, while keeping he oher wo ses of parameers for all he access caegories fixed a he same value (i.e., seing he AIFS values of all access caegories o, which is equivalen o! #"%$'&(!" ). We also elaborae on several implemenaion issues on incorporaing heoreical resuls ino IEEE 80.11e. These include (i) how o reduce he compuaional complexiy and pracically calculae resuls on-line, (ii) how o conver he opimal parameers derived in he model ha characerizes he ) -persisen version of IEEE 80.11e o hose in IEEE 80.11e (which is based on he noion of he conenion window), (iii) how o on-line measure parameers needed for calculaing he bes value of he conenion window size, and and (iv) how o exend he analyic model o incorporae he concep of Transmission Opporuniy (TXOP). Boh he analyical models and he proposed approaches for pracically incorporaing heoreical findings ino IEEE 80.11e EDCA are validaed hrough deailed ns- simulaions and empirical experimenaion on a Linux-based MADWifi driver for wireless LAN devices wih he Aheros chipse. The res of he paper is organized as follows. In Secion II, we give a summary of IEEE 80.11e EDCA operaions ha perain o our work. In Secion III, we presen our analyical model for providing service differeniaion among muliple raffic classes and maximizing he channel uilizaion for IEEE e EDCA. In Secion IV, we exend our analyical model o analyze he role of AIFS and TXOP values on service differeniaion perceived by differen raffic classes. Following ha, we discuss in Secion V several implemenaion issues for pracically incorporaing heoreical resuls ino IEEE 80.11e EDCA, and presen a performance sudy in Secion VI, evaluaing IEEE 80.11e EDCA (wih is parameers on-line uned by he analyical model). Finally, we conclude he paper in Secion VII, wih a lis of research avenues for fuure work. II. PRELIMINARIES In his secion, we summarize he operaions of IEEE 80.11e EDCA ha perain o our work. As EDCA conenion access is an exension of he legacy CSMA/CA DCF mechanism o suppor raffic of differen prioriies, we firs summarize he operaions of IEEE DCF. A. IEEE Disribued Coordinaion Funcion The basic CSMA/CA mechanism in DCF operaes as follows. Each saion mainains a conenion window (CW) and a back-off imer. When a saion has a frame o ransmi, i senses he medium firs. If he medium is sensed idle for more han a ime inerval of Disribued InerFrame Space (DIFS), i sars o ransmi he frame a he beginning of he immediaely following slo. Oherwise, he saion defers is frame ransmission according o he binary exponenial backoff algorihm: The saion wais unil he medium has become idle 3

5 for an inerval of DIFS and hen ses is backoff imer. The backoff imer gives he addiional ime i should wai before he nex ransmission aemp, and is se as he produc of asotime and a pseudorandom ineger ha is drawn from a uniform disribuion over he inerval * +-,.0/. The conenion window is se o is minimum value ( $1-3 ) for he firs ime a frame is backed off. Every ime he frame incurs a collision, is increased o 57683:9<;>=?@3, wih he resricion ha can no exceed is maximum possible value AB (B$C3D+E=F1 ). The backoff imer is decreased by one ime slo for every consecuive idle slo afer he medium has been sensed idle for an inerval of DIFS, and is suspended whenever he medium becomes busy. The backoff process resumes when he wireless medium has been deeced o be idle for an inerval of DIFS. When he backoff imer reaches zero, he saion ransmis he frame immediaely. If he desinaion saion receives he frame correcly, i sends a posiive acknowledgmen (ACK) frame afer a ime inerval of Shor InerFrame Space (SIFS). Noe ha SIFS is shorer han DIFS. When he source saion does no receive an ACK wihin an ACKTimeou inerval, i assumes he frame has experienced a collision, updaes he conenion window according o he binary exponenial backoff algorihm described above, and ses is backoff imer according o a newly seleced random value wihin * +-,.0/. A maximum of 7 reransmissions ( reransmissions) for shor frames (long frames) are allowed before he frame is dropped. The basic access procedure is depiced in Fig. 1 (a). To furher decrease he overhead caused by frame collision and hidden erminal effecs, Reques-To-Send (RTS) and Clear-To-Send (CTS) frames may also be exchanged before he daa ransmission. I is worh noing ha collision avoidance is achieved by a virual carrier sense mechanism. Wheher he channel is idle or busy is no solely deermined by he physical carrier sense resul, bu also by he value of he NAV imer mainained by he MAC. A duraion value is included in each frame ha is ransmied by a saion and indicaes how long he ransmission will las, including any subsequen acknowledgmens and fragmens. Each saion in he viciniy of he ransmiing saion receives he frame and uses he duraion value o updae is NAV. Therefore he NAV value indicaes how long anoher saion has access o he wireless medium no maer wha is he real aciviy of ha saion. B. IEEE 80.11E Enhanced Disribued Channel Access As menioned in Secion I, IEEE 80.11e defines a new HCF coordinaion funcion ha includes boh he conenion-based channel access mehod, EDCA, and he conrolled channel access mehod, HCCA. As EDCA is essenially an exension of he legacy DCF mechanism for which several analyical models exis and can be leveraged, we focus on EDCA in his paper. In EDCA, several parameers ha conrol how and when a node gains access o he medium (e.g., he conenion window, he iner-frame space, and he ransmission opporuniy) differ among differen prioriy levels, so as o favor/disfavor daa ransmission from high-prioriy/low-prioriy flows. A oal of eigh user prioriy levels are available, and mapped o four access caegories (ACs). Each AC corresponds o one of he four ransmi

6 queues ha implemen he EDCA conenion algorihm wih differen parameers. These parameers are he minimum idle delay before conenion (AIFS), he minimum and maximum conenion windows ( and B ), and he ransmission opporuniy limi (TXOP). Specifically, he EDCA funcions under differen ACs wai for differen values of Arbiraion Inerframe Space (AIFS), raher han using he same value of DIFS afer channel has become idle. Fig. 1 (b) depics he relaion beween he various inerframe space (IFS) parameers. Similarly, EDCA associaes differen ACs wih differen values of and, and allows raffic of differen prioriies o back off for differen ime inervals, so as o increase/decrease heir probabiliy of medium access. Finally, a saion ha wins an EDCA conenion is graned a Transmission Opporuniy (TXOP) he righ o use he medium for a period of ime. The duraion of his TXOP is specified per access caegory, and is conained in he TXOP limi field of he access caegory (AC) parameer record in he EDCA parameer se. A saion can use a TXOP o ransmi muliple frames wihin an access caegory. If he DATA-ACK exchange sequence has been compleed, and here is sill ime remaining in he TXOP, he saion may ransmi anoher frame in he same access caegory, provided ha he frame o be ransmied and is necessary acknowledgmen can fi ino he ime remaining in he TXOP. All he parameers can be dynamically updaed by he QoS access poin (QAP) for each access caegory hrough he EDCA parameer se, and are sen from he QAP as par of he beacon frames, and probe/re-associaion response frames. This adjusmen allows saions in he WLAN o adap o changing condiions, and gives he QAP he abiliy o manage he overall QoS performance. When daa frames arrive a he 80.11e MAC layer, hey are classified ino an appropriae AC and enqueued ino he corresponding ransmi queue. When frames are available in muliple ransmi queues for ransmission, conenion for he medium occurs boh inernally and exernally, based on he same coordinaion funcion, so ha inernal scheduling resembles exernal scheduling. Inernal collisions are resolved by allowing he frame wih a higher prioriy o be ransmied, while he flow wih a lower prioriy invokes a queue-specific backoff procedure as if i had incurred collision. III. ANALYTICAL, MULTI-CLASS MODEL The firs sep o analyzing how o une all he parameers in IEEE 80.11e EDCA and supporing service differeniaion is o devise an analyical model ha characerizes he daa ransmission aciviies under EDCA. Our analyical model is buil upon, and exends, ha proposed by Cali e al. in [8, 9, 10]. Essenially Cali s work considers channel access by a single raffic class in a ) -persisen version of IEEE DCF, wih he objecive of improving he sysem channel uilizaion. They do no provide service differeniaion among muliple raffic classes. We assume all saions can hear each oher, and hence here are no hidden erminal and exposed erminal problems. As he major applicaion scenario we consider is an managed wireless LAN in hospos (such as airpors, resaurans and hoels) or home neworks for wireless audio/video disribuion, he assumpion is 5

7 valid. We firs summarize he basic model. Noe ha he basic model is slighly differen from ha in Cali s work, due o he fac ha Cali em e al. assume ha each saion has o wai for an inerval of DIFS afer a frame collision in heir model, while we assume EIFS. 1 Then we elaborae on he proposed muli-class model. The resuls derived in his secion and Secion IV can be used as a guideline o fine-une parameers in real neworks o achieve desired service differeniaion. A. Basic Model For analyical racabiliy, Cali e al. [8, 9, 10] consider a p-persisen version of IEEE DCF, which differs from he sandard proocol only in he selecion of he backoff inerval. Insead of using he binary exponenial backoff imer values, he p-persisen version deermines is backoff inerval by sampling from a geomeric disribuion wih parameer ). Due o he memoryless propery of his geomeric-disribued backoff algorihm, i is more racable o analyze he p-persisen IEEE proocol. The analyic model is derived under he assumpion ha all he saions always have packes ready for ransmission (which is ermed he G-HJILKM)ONQP NSRUT condiion in [8, 9, 10]). Under he geomerically-disribued backoff assumpion, he process ha characerizes he occupancy behavior of he channel (idle slos, collisions, and successful ransmission) ill he end of each successful ransmission is regeneraive, wih he sequence of ime insans corresponding o he compleion of successful ransmission being he regeneraive poins. Cali e al. exploi his regeneraive propery and define he V h virual ransmission ime as he ime inerval beween he jh and (j+1)h successful ransmissions. In such a ime period, idle periods and collisions precede a successful ransmission, where an idle period is a ime inerval in which he channel is idle due o he fac ha all he back-logged saions are in he back-off mode, and a collision is he inerval in which wo or more saions aemp for ransmission and heir packes collide wih one anoher. Le W(XK>9 denoe he average message duraion, and W(ZY\[ 9 he average virual ransmission ime. I hen follows ha he channel uilizaion ] can be expressed as ]M$ W^XK>9 W^ZY [ 9` Le and Ybadc denoe he lenghs of he R h idle period and he R h collision in a virual ransmission ime respecively, e a5fqg he number of collisions in a virual ransmission ime, and " he ime required o complee a successful ransmission. Noe ha " is he ime inerval beween he sar of a successful frame ransmission and he receip of he corresponding ACK plus a DIFS, i.e., "$hki6j"k`"l6mnnop6m&(!"k, 1 As will be elaboraed on laer, i is non rivial o deermine wheher a specific lisening saion will use a DIFS or an EIFS afer a frame collision. (1) 6

8 W { š where K is he ime i akes o ransmi a message. Then, we have su\vxwzy W(ZYq[F9r$ &(!" 6m 6 YbaQc 6j"k`"l6mnno%9 ƒ'6mw@ vxwzy ~ 9 6mW@5" 9 } ~ $ W@ze a5fqg 9 ˆ!5W@ZYua9\6m&(!"l6Š"k!" 6m o%9u6œzwhze a5fqg 9u6'3:9kˆ:W@ -9 6 WhXK>9\6 "k! #" 6mnnoŽ6m&(!", () where he "k`" and o in he firs erm on he righ hand side of Eq.() is due o he exra waiing period in he exended iner-frame space (EIFS) afer deecion of a incorrecly-received frame (i.e., frame collision). Noe ha in Cali s model, i is assumed ha each saion wais for an inerval of DIFS afer a frame collision, while we assume he use of EIFS here. In fac, i is non rivial o deermine wheher a specific lisening saion should use a DIFS or an EIFS afer a frame collision. If he lisening saion is able o synchronize wih one of he colliding frames, and hus iniiae a receiving process, an EIFS will be used. Oherwise (e.g., in he case ha he received power level is comparable for he wo colliding frame preambles), he lisening saion will only perceive a busy channel. In his case, wihou iniializing he recepion of any frame, he saion will use a DIFS. All he simulaion (as well as analyical) models we are aware of do no ake ino accoun his echnical issue, bu simply use eiher DIFS or EIFS exclusively for all he collisions. B. Proposed Analyical Model For racabiliy of analysis, we follow Cali s mehodology and consider a p-persisen version of IEEE 80.11e EDCA. To suppor muliple raffic prioriies, we assume each raffic class use a differen probabiliy, ), o access he channel. For he ime being, we assume ha all he raffic classes use DIFS as he inerframe space value. (We will incorporae he effecs of using differen AIFS and TXOP values on service differeniaion in Secion IV-A.) To faciliae he analysis, we make he following assumpions and noaions: A1) There are classes of saions, each of which conains A) A class-r saion ransmis is frame in a slo (afer he medium becomes idle for DIFS) wih probabiliy ) in he p-persisen version of IEEE 80.11e EDCA (3 R ). A3) The size, K, of a packe sen by a class-r saion is uniformly disribued beween X,d ~ 9, i.e., XK u9 $ + > q, œ žjÿs œ žjÿ ~, 3 > ~ A) All he saions always have a packe ready for ransmission (i.e., he asympoic condiion holds). Recall ha he channel uilizaion in he case of a single class is expressed in Eq. (1). By defining k J dxrq9 $ Š he packe successfully ransmied in a virual ransmission ime is of class-r 9, he uilizaion aained by 7

9 $ $ { Æ $ $ class-r saions can be expressed as ] $ he uilizaion raio beween class-r and class-v as «ª $ ] ] ª W^XK 9ˆ: q J BXR 9 W(ZYq[F9 W^XK 9 :. XRQ9 W^XK ª 9 J B V`9 and he uilizaion raio beween a class-r saion and a class-v saion as «ª $ ]`5 e ] ª e ª eª e ˆ «ª, (3), () To derive ] and ], we need o derive W^ZYu[ 9 and q J BXR 9. As menioned above, in he case of a single class, W^ZYb[ 9 is given in Eq. (). I is easy o see ha Eq. () sill holds in he case of muliple prioriy classes, excep ha W^ze a5fqg 9, W^ZYua.9, and W(zL9 have o be derived. In wha follows, we derive, following he same line of reasoning in [9], W^ze a5fqg 9, W(ZYba.9, and W(zL9 in he case of muliple prioriy classes. Then, we derive q J dxrq9. Derivaion of W(ze auf g 9 : Le e x denoe he number of ransmiing saions in he slo immediaely afer an idle inerval of lengh DIFS, and le a5fqg g f and Ū± a5a5² denoe, respecively, he probabiliy ha a collision occurs and ha a ransmission is successful, boh condiioned on ha a leas one saion ransmis. Then, a5fqg³g f x µ = e x 3:9 $ 3?% Šze x $'+!9k?% Šze x $¹3:9 \½ 3?% Šze u½ x $'+!9 3?ºA» ~ Q3?¼) 9?¾'» } ~ e ) Q3?() 9 u~ bà º ª q Q3? ) ª 9 \½, (6) 3Á? º ~» Q3<?() 9 and ½ ¾'» 5± aua5² x $¹3! e x < 3:9Â$ } ~ e ) Q3<?() 9 u~ À º ª q Q3? ) ª 9 ½ (7) 3?ºA» ~ Q3?¼) 9 The probabiliy disribuion of e a5fqg can hen be expressed as Šze a5fqg $ VL9 $' ê fqg³g f ˆ: kä ± a5au², (8) (5) and W^ze a5fqg 9 as W'ze a5fqg 9Å$ a5fqg $ V`9 ª. u½ 3?º» } ~ Q3? ) 9 ½ ¾'» } ~ e ) Q3? ) 9 u~ º ª q Q3?() ª 9 À?j3 (9) Derivaion of W^ZY a 9 : Recall ha since IEEE does no implemen any collision deecion mechanism, once a collision occurs, i lass unil all he colliding packes have compleed heir ransmissions. The lengh of a collision is hence equal o he maximum lengh of he colliding packes. Specifically, le Ç ª denoe he 8

10 V» $ $ s { s»»»» s {»» 9 lengh of he packe sen by he V h colliding saion, e a he oal number of colliding saions, and eèa ½ he number of colliding class-r saions, hen Yua $'É Ê Ë Ì5Ç ~,BÇ Í,,BÇ \vî:ï DD W^ZY a 9 can be expressed as u½ u½ Ð ½ Ñ { { { WhZYua9r$ ª ª Ð DD ª Ñ *ÒW(ZYba< Fea $ V ~,Be#a Ð $ VJÍ,,Bea Ñ $ V DD ze#a 03:9 ; ze#a $ V ~,Bea Ð $ V:Í,,Bea Ñ $ V ea 83:9U/ DD Under he assumpion of (A3), he condiional expecaion value of Y a, W^ZYua< ea $ V ~,Bea Ð $ VJÍ, 9, can be derived as W(ZYbaµ Fea $ V ~,Be#a Ð $ VJÍ, DD,Be#a Ñ $ V 9Å$ Ó Ô ŠXYuaÁ u9d9 JŸ snõ $ Ó Ô ^?A qf Ÿ \~?A LÖ Ñ ½Ø ª ½ ƒ DD (10) (11),Bea Ñ $ $ ~?jz ~?% :9Ù V 6'3Dƒ@, (1) } ~ and he erm, ze a $jv ~J,Be a Ð $ V Í,,Be a Ñ $jv e a 03:9, can be calculaed as DD ze#a $jv ~,Be#a Ð $ V:Í,,Be#a Ñ $jv e#a 03:9 DD e º } ~» ƒ ) ª ½ ½ Q3? ) 9 `ª ½ V 3? ze x $'+!9k?A ^ze x $¹3:9 e º } ~» ƒ ) ª ½ ½ Q3?¼)O 9 `ª ½ V ½ ½ 3? º» } ~ Q3<? ) 9?¾'» } ~ e ) Q3? ) 9 u~ À º ª q Q3? ) ª 9 By subsiuing Eqs. (1) and (13) ino Eq. (11), W^ZYa9 can be calculaed. (13) Derivaion of W^z-9 : Since a class R saion may ransmi in a slo wih probabiliy )\, we have W^z-9Å$ Nd g f qˆ Æ VˆbÚJ Š5e x j+!9 ˆ`z Š5e x $'+!9B9 ªJÛ ª ~ ½ º» $ Nd g f qˆ } ~ Q3? ) 9 u½ 3<? º } ~» Q3? ) 9 (1) 9

11 $ $ Derivaion of b J BXR 9 : The probabiliy ha he packe successfully ransmied in a virual ransmission ime is of class R can be derived as Finally we have :. BXR 9 $ q J BXR 9 q J B V`9 u½ e ) Q3? ) 9 u~ bà º» ª q Q3? ) ª 9 Ü ¾'» ~ e ) `Q3<? ) 9 u~ À º» ª Q3? )-ª:9 u½ e ) Q3<? ) 9 u~ Ü º» q Q3? ) 9 À eªd)lªeq3? )-ª:9 u~ Ü º» Ùª Q3<? ) 9 e ˆQ) ˆ`Q3? ) ª 9 eª ˆQ)LªÁˆ!d3Á?¼)O 9 By plugging Eqs. (9), (11), and (1) ino Eq. (), one obains he expression of ] in he muliple prioriy class case. Deerminaion of ) values ha maximize he channel uilizaion Wihou loss of generaliy, we express all he flow hroughpu requiremens in erms of he relaive raio o a class 3 flow (i.e., X~ ). For clariy of presenaion, we also assume ha he daa frame size of all raffic classes are of he same disribuion, ha is, Ý RB, ÝÙVF, WhzK 9 $'WhXK ª 9, hen we can express ) as a funcion of ) ~ using Eq. () and Eq. (16), i.e., ) $ Z~ ˆQ) ~ X~ ˆd) ~ 6ŒQ3<? ) ~ 9` Now he proocol capaciy can be opimized by finding he opimal value of ) ~ ha maximizes ] (Eq. (1)) subjec o he consrain on he relaion beween ) an ) ~ (Eq. (17)). In Secion V-B, we will show how o conver he opimal value of ) in he ) -persisen version o he opimal value of he conenion window in he conenion-window-based EDCA. (15) (16) (17) IV. INCORPORATING THE EFFECT OF AIFS AND TXOP INTO THE ANALYTICAL MODEL A. Incorporaing AIFS in he Analyical Model In Secion III, we assume afer a busy period, all he saions have o wai for he channel o become idle for a ime inerval of DIFS before hey sar o decrease heir backoff imers. Recall ha in he IEEE 80.11e draf, differen AIFS values can be assigned o differen raffic classes, which represens anoher dimension of design freedom. Table I gives he defaul values of, B and AIFSN for differen raffic classes. Noe ha for access caegory R, is!" value µ`"á* R / is deermined by is!"ke value µ`"kem* R /, ha is, µ`"á* R /u$0"k! #"È6!#"ke * R5/-;MGÙ"kÞzP NSY<R5K ß. Alhough i is inuiive o assign larger AIFS values o low-prioriy raffic for service differeniaion, i is no clear o wha exen exending he design space along his dimension faciliaes service differeniaion. Moreover, he defaul values specified in he draf sandard (Table I) are no backed up by heoreic analysis. In his secion, we exend our analysis model in Secion III and sudy he impac of differen AIFS values on he channel hroughpu aained by flows of differen raffic classes. 10

12 $ š $ š ä á Í á ˆ á ä Í For ease of analysis, we consider only wo raffic classes. As i is likely ha he AIFS values are only used o differeniae low prioriy background raffic from high prioriy raffic, an analysis wih wo raffic classes shed lighs on how AIFS values affec QoS differeniaion. Moreover, he model can be sraighforwardly exended o he case of muliple raffic classes. To faciliae he analysis, we make he following assumpions and noaions: A1 ) There are wo classes of saions, each of which conains e saions (3 R Š= ). A ) A class-r saion aemps o ransmi a frame in an idle slo wih probabiliy ) in he p-persisen version of IEEE 80.11e EDCA (3 R Š= ). A3 ) All packes are of he same size Ç. A ) All he saions always have a packe ready for ransmission (i.e., he asympoic condiion holds). A5 ) Insead of waiing DIFS a he end of each busy channel period, a saion of class-r wais for!"á* R5/ ime, where µ`"á* R / $"k`"a6'gù"kþzp NSY<R5K ß ;Aµ`"kem* R /, before i aemps o access he channel wih probabiliy )q a he beginning of each subsequen idle ). For ease of noaion, we denoe µ`"kem*3 / and!"kem*ò= / as ~ and µí respecively. Wihou loss of generaliy, we assume ~ jí. Figure depics an idle period beween wo channel busy periods. Le denoe he number of idle slos afer SIFS ime bu before any saion sars o ransmi. The probabiliy mass funcion of can be expressed M$ VF/ +-, +È V Š~J, B3Á? )\~.9 Sâ ªJ Oãb 3Á?ŠQ3<? )\~.9 Sâ, ~ V^ j Í, Ü d3á?¼) 9 â ªJ Oã Ü Ü.æ ˆ å.3? Q3?¼) 9,ç Í V ~ ~ The condiional probabiliy ha he busy slo ha follows he V idle slos incurs a collision (no collision) can be derived as Š collision M$jV`9 success M$ V`9, (19) success #$jv`9 +-, + V^ ~, e ~ ) ~ Q3<? ) ~ 9 d u~ 3<?ŠQ3? )\~.9, ~ V^ jí, e ~ ) ~\Q3<? )\~.9 d u~ Ð Ð B3Á? ) Í 9 6 e Í ) Í Q3?¼) Í 9 u~ d3á?¼)\~ 9 3Á?@d3Á? ) ~ 9 Ð,çµÍn V ˆ`Q3? ) ÍJ9 The probabiliy ha a class-3 or class-= daa frame is successfully ransmied under he condiion #$jv can (18) (0) 11

13 { { $ { Æ { ˆ s { š š ˆ be derived as and ŠzTèÞzGLH H ~ success È$ V`9 $ ŠzTèÞzGLH H Í success È$ V`9 $ +-, +È V^ j ~, e ~ ) ~ Q3? ) ~ 9 u~ 3?ŠQ3?¼)\~.9, ~ V ŠµÍ, e ~z)\~q3? )\~.9 Q u~ Ð Q3? ) Í 9 3Á?Šd3Á? ) ~ 9 Ð, Í VF, ˆ!Q3?¼) ÍJ9 +-, +È V^ j ~, +-, Ð ~ V ŠµÍ, e#ís) Í Q3? ) ÍJ9 u~ Q3? ) ~ 9 3Á?Šd3Á? ) ~ 9 Ð, Í V ˆ!Q3?¼) ÍJ9 Le he probabiliy ha a successfully ransmied daa frame is a class-3 or class-= frame be denoed as Š* TèÞzGLH H ~ success/ and Š*ÒTèÞzG-H H Í success/ respecively. They can be derived as follows: H ~ success9å$ Š TèÞzG-H H Í #$ VL9 ˆ: ŠzTèÞzG-H H ~ success È$ VL9 ã Ð u~ Ú`Q3?¼) ~ 9 dû ªJ Oãb Ĵe ~ ) ~ d3á? ) ~ 9 u~ ª bãu 6 Æ ª bã Ð Ú`Q3? ) ~ 9 Û ªD Oã Ð ˆbÚ`Q3? )qíd9 Û ªD Oã Ð Ĵe ~ ) ~ Q3? ) ~ 9 d u~ Ð d3á? ) ÍJ9 $ e ~ ) ~ Q3? ) ~ 9 u~ Ð ˆ`Q3?() Í:9 3?ŠQ3? ) ~9 œ ã Ð Oãb U 3?ŠQ3? ) ~ 9 6 Q3?¼)\~.9 œ ã Ð Oãu 5 Ð Q3<?() Í 9 3Á?Šd3Á? ) ~ 9 Ð ƒ, (3) ˆ!Q3? )qíd9 $ ÞzG-H H Í success M$ VL9 ª $ Æ ª bã Ð Ú!Q3? ) ~ 9 BÛ ªD Oãu Ð ûú!q3? )qíd9 Û ªD Oã Ð Ð Ĵe#Í ) Í d3á? ) ÍJ9 u~ Ð $ e Í ) Í Q3? ) Í 9 u~ ˆ`Q3?¼)\~.9 Q3? ) ~ 9 œ ã Ð Oãb U 3?@Q3? ) ~ 9 Ð ˆ!d3Á? ) Í:9 Q3? ) ~ 9 (1) () () 1

14 $ é $ á ˆ á ˆ á ˆ ˆ á æ The raio,, of he average flow hroughpu of he wo raffic classes can be derived as $Žé ~e HE~ success9le Í ÍDe H Í success9le ~ e ~ ) ~ Q3? ) ~ 9 u~ ˆå 3<?ŠQ3?() ~ 9 œ ã Ð Oã 3<?ŠQ3? )\~.9 6 Q3? ) ~ 9 œ ã Ð Oã Ð ˆ`Q3?¼) ÍJ9 3Á?@d3Á? )\~.9 Ð eí ˆ`Q3? ) Í 9 Ð e Í ) Í Q3<? ) Í 9 u~ ˆLQ3?()\~ 9 Q3<? ) ~ 9 œ ã Ð Oãu 5 3?ŠQ3? ) ~ 9 Ð e ~ ˆ!d3Á? )qíd9 $ 3Á?@d3Á? ) ~ 9 Ð ˆ!Q3? ) ÍJ9 â 3Á?ŠQ3<? ) ~ 9 œ ã Ð Oã Xâ ê Í ~ ˆµëìè3 6 Ð Q3<? ) ÍJ9 3Á?ŠB3Á? ) ~ 9 â Q3? ) ~ 9 œ ã Ð Oã Xâ íî, (5) where é ~ and é Í are he aggregaed hroughpu of class- 3 and class-= raffic respecively, and ê Í ~ $ )\~ d3á? ) Í 9 is he raio of he average flow hroughpu of he wo classes in he case ha flows of wo ) Í d3á? ) ~ 9 differen classes use he same µ`" value, bu differen ransmission probabiliies ) ~ and ) Í respecively. An imporan implicaion can be made on Eq. (5) he raio of average per flow hroughpu beween differen raffic classes is a funcion of boh he ransmission probabiliy ) (or equivalenly he conenion window size) and he AIFS values?l X u~. As he number of prioriy classes increases, i becomes increasingly difficul o adjus unable parameers )q and µ-? <X u~ (3 R Š ), in order o simulaneously mee all consrains. (This will be corroboraed by he simulaion sudy in Secion VI-A.0.d.) I would acually be desirable o decrease he dimension of design freedom by eiher (i) seing he AIFS values of all he access caegories o he same value (i.e., ) o enforce! #"%$'&(!" or (ii) seing he values of ) o pre-deermined values and deermining he appropriae values of o mee he consrains (Eq. (5)). The former approach is wha has been aken in Secion III, where i has been shown ha hroughpu differeniaion requiremens can be me by adjusing ) while keeping all values of AIFS of all access caegories he same. In he laer approach, all he analysis performed in Secion III remains unchanged, excep ha (a) he DIFS erm in Eq. () is replaced by "k`" ; and W^z-9 is calculaed using Eqs. (18) (). One imporan observaion is ha ~ neiher conribues o hroughpu differeniaion nor improves he channel uilizaion. Therefore i should be se a small value (i.e., which is equivalen o! #"k~ $'&(!" ). B. Incorporaing Transmission Opporuniy (TXOP) in he Analyical Model In addiion o he minimum idle delay before conenion (AIFS) and he minimum and maximum conenion windows ( and B ), anoher imporan parameer ha is associaed wih each access caegory in 80.11e is he Transmission Opporuniy (TXOP). In he legacy DCF proocol, a saion can only ransmi one MAC Proocol Daa Uni (MPDU) for each successful conenion. In EDCA, a TXOP is defined by a saring ime and a maximum duraion. Afer a saion successfully conends for he channel, i is allowed o ransmi muliple MPDUs as long as he ransmission ime does no exceed he TXOP limi. As depiced in Fig. 3, wihin a TXOP, he TXOP holder saion coninuously ransmis he nex MPDU, a "k!" ime inerval afer i receives he o frame corresponding o he previous MPDU. 13

15 { To incorporae he effec of TXOP ino he derived analyical model, we noe ha in EDCA wih TXOP, only he firs MPDU wihin each TXOP burs may collide wih MPDUs from some oher saions. Tha is, as far as he analyical model is concerned, we can envision each TXOP burs as an exended frame ransmission. Therefore, W(ZYua.9 in Eq. () can be calculaed wih he use of he disribuion of he ïð & ñ size, and " in Eq. () can be calculaed as he TXOP limi minus he residual ime a he end of a TXOP (which canno accommodae one more MPDU and is corresponding ACK and is noed as he wase ime in Fig. 3). where oö$ø xy<ùú K ó»\ô õ "$'&^`"l6moòˆ!zk ó»\ô õ 6j"k`"l6mnnop6 "k`" 9k? "k!"k, (6) limi èï8 & ñµû is he number of MPDUs ha can be accommodaed in a TXOP, is he ransmission duraion of a MPDU, and W(ZYq[ 9 $'W s \vxwzy } ~ z&(!" 6m 6 YbaQc 6j"k`"l6mnno%9 ƒ 6mWhz vxwzy ~ 9 6mW@U" 9 Noe ha under he asympoic scenario, wheher or no a saion successfully conends for channel access is independen of wheher or no TXOP is used, and hence he derivaion of he probabiliies of collision and successful ransmission is sill valid. Following a similar derivaion as in Secion IV-A, one can analyze he effec of TXOP limis among differen raffic classes on service differeniaion. Foreseeable, a similar conclusion may be drawn, i.e., wih muliple dimensions of conrol knobs (, B, AIFS, and TXOP), i may become inracable o simulaneously mee all he proporional hroughpu consrains among differen classes. How o fine-une TXOP policies/parameers under he derived analyical model wih he oher parameers fixed is a subjec of our fuure sudy. V. INCORPORATING THEORETICAL RESULTS INTO IEEE 80.11E EDCA There are several implemenaion issues ha we mus address in order o incorporae he heoreical resuls derived in Secion III ino he operaions of IEEE 80.11e EDCA. Firs, we need o devise an approximae soluion o calculae he opimal values of ) wih reasonably small compuaional complexiy and error discrepancy (due o approximaion). Second, we need o associae he derived opimal values of ) in he ) -persisen version wih he conenion window based backoff scheme in EDCA. Tha is, we have o find a mapping beween he conenion window, T.ü, and he persisen probabiliy, ). Third, as he opimal values of he sysem parameers (i.e., ) or equivalenly ) change wih he raffic densiy, we need a simple ye effecive mechanism o on-line esimae he number of acive saions in a QoS basic service se (QBSS) so as o opimize he sysem performance in he presence of nework dynamics. (7) 1

16 s ä» & ý {» ä» s s ä» {» Ö {» 6 {» 6 Õ ˆ 6 Y Y A. An Approximae Soluion o Obaining he Opimal Value of ) To faciliae he derivaion hereafer, we define he following variables 0ý } ~ u½ Q3? ) 9, þÿý ~ e ˆQ) 3Á? ) Ö, (8) ðý'w^zyua9\6m&(!" 6j"k! #"l6mnno>, (9) } ~ e X~, ý ~ e Í Z~, Y'ýhNd g f (30) W@5e a5fqg 9 and Wh L9 can hen be expressed in erms of and þ as W^ze a5fqg 9 $ (31) and W^z-9 $@YŠˆ 3? To opimize he channel hroughpu, we only need o minimize W@zY [ 9 (Eq. ()). Afer omiing he consan iems in Eq. (), he problem of finding he opimal value of ) ha maximizes he channel uilizaion reduces o one ha minimizes As ) $ X~ ˆQ)\~ X~ ˆd) ~ 6ŒQ3? ) ~ 9 Q3?%n9ˆ hĵþ (Eq. (17)), B can be rewrien as þÿ$ Thus he problem is equivalen o minimizing By defining $ minimizes ) ~ 3? ) ~ } ~ } ~ e X~ ƒ Õ ½ X~ ˆB) ~ 6ŒQ3? ) ~ 9 Q3?¼) ~ 9 Ú ¾ } ~» e X~ Û )\~ 3? ) ~ To furher simplify Eq. (36), we noe ha em )\~ 3?¼) ~?j3 ƒ ˆ Y þ $'& )\~ 3Á? ) ~ Ú ¾ } ~» e Z~ Û )\~ 3? ) ~, he problem can be recas as one ha finds he opimal value of 8X +!9 ha } ~ Q36 u½ X~ b9?j3 ƒ ˆ (3) (33) (3) (35) (36) X~d 3. This is based on he observaion ha under normal, non-conenion condiions, even if only class R nodes are acive, he probabiliy ha a leas one saion 15

17 as ä» {» œ ô Ð Õ = Í = œ ô Ð Í sars o ransmi a he beginning of an idle slo is far less han 1 (oherwise conenion occurs), and hence e X~ ¼$'e ½ ~d ½ 3. Wih e X~ 3, we can make he following approximaion: d3 6 u½ X~ b9 36 Ú ¾ } ~» X~ e Û Í?¹Ú ¾ } ~» X~ e È6 } ~ ~ $ 36m&iˆD M6 & Í?% = ˆD and hence Eq. (36) can be rewrien as As & Í?% &iˆ Š6 j+, Eq. (38) is minimized when & Í? = Ö ˆ ŒˆD M6 ¼$ Í z& Í?A9 ˆ = Y \ When 3, we have ) ~ $ ~ >$ Ã. Tha is, we use Í of ) ~ ha opimizes he channel uilizaion. B. Mapping he p-persisen Probabiliy o he Conenion Window Size X~ Í e Û, (37) Y M Ĵ (38) \ à o approximae he value As discussed in Secion II, EDCA in he curren 80.11e proocol employs a conenion window based backoff mechanism. If we assume he backoff couner value XNB9 is randomly chosen in * +-,.0/, and he conenion window is fixed (hroughou he duraion in which a saion aemps o ransmi a frame), he probabiliy ha a saion ransmis in a slo can be derived (wih he use of he same echnique in [5]) as )¼$ $p = T ü 6 = Eq. (0) enables us o apply he analyical resuls in Secion III o he conenion-window-based EDCA proocol, by seing and B o ) )?= where is he opimal ransmission probabiliy calculaed using our analyical model in Secion III. Tha is, wih he conenion window size, he probabiliy ha a saion ransmis in a slo is equal o he opimal ransmission probabiliy derived in our analyical model. C. On-line Measuremen of Parameers Needed for he Analyical Model (39) (0) + ûf, (1) In he muliple-class, ) -persisen analyical model, he number of saions compeing for he channel is assumed o be fixed and known a ) priori. In pracice, he number of back-logged saions varies, which necessiaes esimaion of he number of back-logged saions ha compee for he wireless medium, so ha he opimal ransmission probabiliy can be deermined accordingly. In his subsecion, we devise a simple 16

18 $ { ª { and ye effecive measuremen mechanism o on-line esimae he sysem parameers (i.e., he number, e, of back-logged saions in each access caegory) ha are required o calculae he opimal values of ) (and ). In he case ha here exiss one prioriy class, where all saions use he same ransmission probabiliy, he average idle period duraion beween wo consecuive ransmissions, W'ZY g ² 9, can be expressed as (Eq. (1)) where) is he ransmission probabiliy and e W@zY g ² 9 $ Q3? )u9 ˆDNQ g³f Q, () 3?@Q3? )b9 is he number of acive saions. When he ) -persisen probabiliy is known, one can infer he number of acive saions by on-line measuring he idle period. In he case of muliple prioriy classes, alhough he average idle period duraion beween wo consecuive ransmissions can sill be expressed as a funcion of he number of saions in each class and he ransmission probabiliy of each class, i.e., Eq. (1), i is difficul o use his relaion o esimae he number of acive saions in each class and o ensure he uniqueness of he soluion:eèq,j3 jr j. To deal wih he problem, we propose o keep rack of he number of acive, class-r saions from he channel access hisory overheard in he pas successful ransmissions. The value of deermines he rade-off beween he accuracy and he sensiiviy of he online esimaion algorihm. The larger he value of, he smaller he probabiliy of missing saions ha experience large access delays due o muliple consecuive collisions. On he oher hand, a large value of^ also implies a slower response o saion sae changes. We seè o be he value such ha he probabiliy ha any given, class-r saion successfully ransmis a leas one daa frame is larger han a predefined hreshold (e.g.,$œ+ As he probabiliy ha a given saion ). successfully ransmis in a virual ransmission ime can be calculaed as XR 9ø$ \½ ) Q3?¼) 9 u~ bà º» ª q Q3?¼) ª 9 u½ ¾ ~» e ) Q3<? ) 9 u~ bà ˆ º ª» q Q3?¼) ª 9 u~ )» eª ˆQ)Lª 3<? )O ª ~ 3? )Lª we can se he value of o be he larges ineger such ha ª ~ z zr 9d9 d3á?% XR 99 `ª, (3) Eq. () ensures ha in he esimaion period, he probabiliy ha any given acive, class-r saion successfully ransmis a leas one daa frame is larger han. D. Complee Procedures for Realizing Service Differeniaion in IEEE 80.11e EDCA By incorporaing he above (i) on-line algorihm for esimaing he number of acive, class-r saions and (ii) mapping beween he opimal ) -persisen probabiliy and he conenion window size, EDCA will be 17 ()

19 able o suppor (proporional) service differeniaion. The complee procedures ha he AP and each of he mobile saions ake are described in Fig.. Noe ha in he procedures, raher han having boh c and c B se o (as suggesed in Secion V-B), only c is se o he opimal value calculaed according o he analyical model. Tha is, boh he AP and mobile saions sill carry ou he exponenial binary backoff algorihm for medium access conrol, excep ha c $p. This enhances he robusness of he proocol under some abnormal scenarios, such as he emporary inerference and deviaion of he esimaed number of acive saions from is acual value. Under hese cases, frame collision sill occurs and he exponenial binary backoff algorihm will ake effec o miigae collision. On he oher hand, under normal cases, collision does no occur frequenly when he parameers are updaed wih quasi-opimal values, and hence he performance discrepancy beween hese wo choices should no be noable. VI. PERFORMANCE EVALUATION To validae he correcness of our analyic model and o evaluae he performance of IEEE 80.11e EDCA wih he analyical resuls incorporaed, we have implemened he p-persisen version of 80.11e EDCA in ns-. We have also implemened an experimenal prooype of he enhanced EDCA mechanism (ha includes he service differeniaion algorihm described in Secion V-D) on a Linux-based MADWifi driver for wireless LAN devices wih he Aheros chipse. In wha follows, we repor our simulaion and empirical resuls. A. Simulaion Sudy In he simulaion sudy, he nework opology includes e class-r mobile saions, each of which has a CBR raffic source ha generaes packes eiher (i) a a rae high enough o emulae he asympoic condiion or (ii) in compliance wih he on-off model o emulae bursy raffic. All saions send CBR packes of sizes 500 byes o he base saion. We do no consider TCP raffic, because we are primarily ineresed in he performance of IEEE EDCA and do no inend o model he ineracion of TCP wih EDCA. (The reason for no modeling he ineracion is in par due o he difficuly and complexiy involved in characerizing he inrinsic characerisics of TCP in addiion o MAC proocols hemselves). In realiy i is also unlikely QoS sensiive raffic will be ranspored hrough TCP. The values of sysem parameers used in he simulaion conform o he 80.11b sandard, and are lised in Table II. The opimal class-r ransmission probabiliy, )~, is calculaed using numerical ) mehods (Secion III- B) or using he approximae soluion (Secion V-A), wih he objecive of maximizing he channel uilizaion (subjec o he hroughpu raio consrain). The opimal ransmission probabiliies, XR =E9, are calculaed using Eq.(17). We calculae he hroughpu in erms of he received payload of MAC daa frames, and carefully consider all he overhead inroduced in he physical and MAC headers in boh he heoreic analysis and he simulaion. We assume ha all he saions can hear each oher, and hence do no consider he hidden erminal/exposed 18

20 erminal problem. We presen only simulaion resuls in he -prioriy class case. Unless specified oherwise, he argeed raio, ê Í ~, of he hroughpu aained by a class 1 saion and ha by a class saion is se o.0. Simulaions resuls under oher nework configuraions can be found in [1]. a) Validaion of he Analyic Model: In his se of simulaion, we validae he analyic model in Secion III wih respec o is capabiliy of providing hroughpu differeniaion and achieving he maximum channel uilizaion. For he purpose of comparison, in addiion o using he opimal ) ~ and )qí values, we also!#" use 18 ses of ransmission probabiliies o calculae he channel uilizaion, where ) ~ is seleced from d)~ ;l+ 3;¼R 3 R andb)~ ;d3 +µ6m+ R 9\ 3 R 03D+$ and ) Í is calculaed using Eq.(17). Figures 5 depic he oal sysem hroughpu, he hroughpu aained by each class, and he per flow hroughpu raio in he case of e ~ $ÿ3&% and e#í$'. Several observaions are in order: Firs, he simulaion resuls are in exremely good agreemen wih hose obained in he analyic model. Second, when he ransmission probabiliies deviae from he opimal value, he oal sysem hroughpu decreases accordingly, as prediced in he analyical model. Also, he hroughpu raio beween he wo classes is very close o he specified value, indicaing ha QoS provisioning hrough appropriae seing of backoff values is feasible. This is also in line wih our analysis. b) Validaion of he Approximae Soluion: To validae wheher or no he approximae soluion proposed in Secion V-A (Eq. (39)) renders accepable resuls, we conduc simulaion o compare he hroughpu resuls wih he opimal and approximae values of ) ~ respecively, by varying he values of ~ Í, he sizes of daa frames, and he number of saions in each class. Again, due o he page limi, we presen only one se of simulaion resuls in Table III where he packe size is se o 500Byes. As given in Table III (a), he sysem hroughpu obained wih he approximae values of ) comes surprisingly close o ha wih he opimal value of ). To beer undersand why such he approximae soluion renders such good resuls, we lis in Table III (b) he value of (Eq. (9)) which corresponds o he fixed porion of W^ZYb[ 9. In fac, is equal o he value of W(ZY\[ 9 assuming perfec scheduling algorihm is employed. As boh he opimal value and he approximae value of ) ~ resul in very low frame collision raes, in he viciniy of he opimal nework operaion sae, he frame collision rae is quie small and he overhead incurred in frame collision and binary backoff only couns for a small porion of W(XN[ 9. Even if he approximae value of ) ~ canno achieve he minimal overhead duraion (i.e., W^ze a5fqg 9 ˆ.%6 zw(ze a5fqg 9Ù6 3:9 ˆBW^zL9 ), he aggregaed channel hroughpu is sill close o is maximal achievable value. c) Performance of IEEE 80.11e EDCA wih Analyical Resuls Incorporaed: To evaluae he performance of IEEE 80.11e EDCA wih he analyical resuls incorporaed (Secion V-D), we have carried ou simulaion under various scenarios. Single Class, Greedy Traffic: In his se of simulaions, we evaluae he performance of IEEE 80.11e EDCA wih he on-line parameer esimaion mechanism incorporaed (Fig. ) in he case of a single raffic class. 19

21 There are a oal of 10 mobile nodes and one access poin. Each mobile node generaes daa packes wih a rae high enough o saurae he channel (i.e., he asympoic scenario). The average daa packe size is 500 byes. We acivae one saion every five seconds, and 30 seconds afer all saions are acivaed, we sar o deacivae one saion every five seconds. We keep rack of he number of acive nodes (boh acual and esimaed), he calculaed ransmission probabiliy, ) ~, and he sysem hroughpu every seconds and depic hem in Fig. 6. The simulaion resuls show ha he number of acive nodes esimaion algorihm works quie well and he sysem hroughpu is kep high due o he fac ha he ransmission probabiliy is adapively calculaed o miigae collision. Two Classes, Greedy Traffic: In his se of simulaions, we evaluae he performance of IEEE 80.11e EDCA wih he analyical resuls incorporaed in he case of wo raffic classes. The raio of per flow rae ê ~ Í is se o.0. A oal of 0 mobile nodes exis in he WLAN, wih 10 nodes in each class. Each node generaes daa packes wih a rae high enough o saurae he channel (i.e., he asympoic scenario). We acivae one saion in each class every 10 seconds, and 50 seconds afer all saions are acivaed, we sar o deacivae one saion in each class every 10 seconds. As shown in Fig. 7, for mos of he ime, he raio of per-flow aainable hroughpu beween he wo classes is very close o ê ~ Í, i.e., he sysem capaciy is disribued among differen raffic classes in compliance wih he hroughpu requiremen. Moreover, he channel uilizaion is kep high regardless of he change in he number of acive nodes. Two Classes, On-Off Traffic: In his se of simulaions, we sudy he performance of he analyically enhanced version of IEEE 80.11e EDCA in he case of bursy raffic. The nework configuraion is similar o ha in he previous se of simulaions, excep ha each mobile node generaes on-off raffic. The duraion of he on and off periods are boh exponenially disribued wih mean ime 5000ms. The daa packe generaion rae in he on period is se o 500K bps. The daa packe size is 500 byes. As shown in Fig. 8, he raio of per-flow aainable hroughpu beween he wo classes is kep, for mos of he ime, a he desired value (i.e.,.0), alhough he per flow hroughpu raio is no sricly achieved when he channel offered load is low (in he firs one hird and he las one hird simulaion period). The failure o keeping he desired hroughpu differeniaion during any ime inerval in he case of on-off raffic is, in par, due o he fac ha he analyical resul is derived under he fluid model assumpion, which does no hold in he case of bursy raffic. The same phenomenon has also been observed and discussed in [16]. d) Effec of AIFS on Service Differeniaion: As menioned in Secion IV-A, he raio of average per flow hroughpu beween differen raffic classes is a funcion of boh he ransmission probabiliy ) (or equivalenly he conenion window size ) and he AIFS values?j Z u~ (Eq. (5)), and i becomes difficul o simulaneously une boh ses of parameers in order o mee all he hroughpu raio consrains, while maximizing he channel uilizaion. To validae he above saemen, we carry ou simulaion wih differen AIFSN values. In his se of simulaions, wo groups of mobile saions exis in a BSS, each of which conains K and( 0