Fair Scheduling for Real-Time Multimedia Support in IEEE Wireless Access Networks

Transcription

1 Far Schedulng for Real-Tme Multmeda Support n IEEE Wreless Access Networs Yaser P. Fallah Inst. of Transportaton Studes (Depts EECS & CEE), Unversty of Calforna at Bereley, USA e-mal: yaserpf@eecs.bereley.edu Panos Nasopoulos Dept. of Electrcal and Computer Eng. Unversty of Brtsh Columba Vancouver, Canada e-mal: panosn@ece.ubc.ca Raja Sengupta Inst. of Transportaton Studes (Dept of CEE), Unversty of Calforna at Bereley, USA e-mal: sengupta@ce.bereley.edu Abstract - Successful deployment of Broadband Wreless Access Networs such as WMAX (IEEE ) wll be contngent on provsons for supportng multmeda traffc. In ths paper, we revew the qualty of servce features of access networs such as the standard, and dentfy algorthms and schemes that are needed for supportng multmeda traffc n such networs. The standard only specfes the features that should be mplemented and leaves the desgn of a qualty of servce soluton to developers. Ths ncludes the desgn of a mandatory schedulng framewor. We present a comprehensve multmeda support framewor based on the standard features. The framewor specfes the archtectures for the base staton and the subscrber staton, and contrbutes a number of algorthms for dfferent servce provsonng objectves. We use the concept of vrtual pacets to provde far pacet based centralzed schedulng of upln and downln pacets. The presented soluton also provdes algorthms for temporal and throughput far schedulng n multrate physcal layer of the networs. An mportant part of the presented desgn s a mult-class far schedulng scheme whch s proposed for provdng better delay performance for real tme applcatons, whle mantanng slghtly longer term farness. Keywords - WMAX, QoS, Schedulng, SFQ, Multmeda, IEEE , MAC, TDMA. I. INTRODUCTION IEEE s the recently approved standard for wreless metropoltan area networs (WMAN) [1] [2]. Ths standard, whose ndustry-led equvalent s also referred to as WMAX, wll be a compettor to the other currently avalable broadband solutons such as DSL, cable, and IEEE ( [3]) n offerng broadband access. Thus, t s necessary to develop solutons for supportng popular qualty of servce (QoS) demandng applcatons such as real tme multmeda n WMAX networs. Whle QoS support has been consdered durng the development of , the standard only descrbes the requred features and classes of QoS and does not defne the desgn of a mandatory scheduler used for provdng the requred QoS. Such a desgn s left to vendors of WMAX base staton and subscrber statons. Devsng schedulng algorthms for WMANs s a challengng tas for the followng three reasons. Frst, defnes several classes of traffc wth dfferent types of QoS requrements (mult-class QoS); conventonal far schedulng or prorty based schedulng cannot acheve both far throughput guarantee and the desred dfferentated delay performance at the same tme. Second, statons may use dfferent transmsson rates at dfferent tmes (multrate operaton). And thrd, upln and downln traffc n , though separated usng duplexng methods, ultmately share the same tme resource (assumng that tme dvson duplexng, TDD, mode s used n the MAC, as opposed to the less effcent and less flexble frequency dvson duplexng, FDD). Our proposed soluton, n ths artcle, addresses these ssues and effcently utlzes the avalable features of the standard to provde far guaranteed servces for real-tme applcatons. In the past few years, several QoS solutons for networs have been developed. However, to the best of our nowledge no model s provded for far mult-class servce provsonng and smultaneous schedulng of upln and downln pacets n an MAC. Hawa and Petr presented an archtecture for schedulng upln and downln flows separately [4]. Ccconett et al. have conducted a performance study to evaluate QoS provsonng avalable n the MAC layer, operatng n the FDD (frequency dvson multplexng) mode [5]. In another wor, Cho et al. proposed a new MAC archtecture wth a bandwdth allocaton, as well as an admsson control technque for better QoS support n [6]. Gusa et al. carred out a comparson of two well-nown schedulng schemes namely, WFQ and WRR when used n MAC n terms of average pacet delay [7]. They also proposed an algorthm for dynamc rato adjustment of UL and duraton wthn a frame. Although the exstng solutons can enhance the performance of the MAC n QoS provsonng, they do not address the farness ssues and the combned schedulng of upln and downln flows (necessary for the TDD mode). Moreover, the multrate operaton of the physcal layer (PHY) s not consdered n these wors. In ths paper, we present a detaled archtecture and ts supportng algorthms for provdng far mult-class servces for upln and downln flows n an MAC, consderng the multrate operaton. The proposed archtecture also allows for other algorthms to be used as ts components. We note that provdng farness s necessary n cases that customers expect to receve a far servce accordng to ther contract. Farness becomes especally mportant n the case of multrate operaton of the PHY, due to the fact that one staton s poor channel may degrade servce qualty for others. In the next secton some basc techncal detals of the MAC layer of the IEEE standard are descrbed. Secton III presents a detaled descrpton of the proposed archtectural

2 desgn followed by the performance evaluaton of the model dscussed n Secton IV. Fnally, we conclude the artcle n Secton V. II. OVERVIEW OF THE IEEE STANDARD The MAC s connecton orented; each subscrber statons (SS) establshes a connecton wth the base staton (BS) to delver voce, vdeo, or usng connectons wth dfferent QoS requrements [1]. In contrast to the standard [3], acheves duplex operaton through ether FDD or TDD. The downln channel (from BS to SS) s a broadcast channel, whle the upln channel (from SS to BS) s shared by SSs. Tme n the upln channel s usually slotted (mn-slots) and shared usng tme dvson multple access (TDMA), whereas on the downln channel BS uses a contnuous tme-dvson multplexng (TDM) scheme. Other multple access schemes such as orthogonal frequency dvson multple access (OFDMA) have also been specfed n later versons of the standard [2]. The duraton of the downln or upln subframe n TDD mode s determned by the BS n a dynamc manner and s based on the amount of upln or downln load. The MAC layer requres that the BS and SSs be synchronzed at all tmes and transmt only n predetermned tme slots or subcarrers. In TDMA mode, the upln channel s dvded nto a sequence of mn-slots, whch SSs can access n a synchronzed manner controlled by the BS. If OFDMA s used, subcarrers are dvded amongst SSs to allow smultaneous transmsson n the same slots. The BS uses Upln and Downln maps (UL-MAP and -MAP) at the begnnng of a subframe to ndcate the assgnment of mnslots (e.g. Fg. 1) or subcarrers. TDD freq. FDD Map Map UL Map UL Map Downln Subframe Contenton UL slots Scheduled UL Contenton UL slots Scheduled UL Upln Subframe Scheduled UL Fg. 1. IEEE MAC Frames Scheduled UL Gven that transmsson s always controlled by the BS, schedulng becomes one of the central tass of the MAC. The new revson of the standard (802.16e) revses the descrpton of QoS servces offered and n partcular clarfes the dstncton between upln and downln flows [2]. Fve types of upln servces are defned: Unsolcted Grant Servce (UGS); Extended Real-Tme Pollng Servce Real-Tme Pollng Servce (rtps); Non-Real-Tme Pollng Servce (nrtps) and Best Effort (BE) servce. However, QoS descrptors used tme tme for specfyng these upln servces are also used for downln servces, and the dstncton s only n the modes of BW request for upln flows. Downln pacets can be drectly scheduled by the BS n the downln subframe, accordng to the QoS descrpton of ther connectons. For upln connectons, each SS that needs to send has to frst request bandwdth from the BS. The BS wll then assgn upln transmsson opportuntes to the SS n the next upln subframes. Two modes of bandwdth (BW) requests are defned: contenton based and pollng based. In contenton mode, SSs send BW requests durng predetermned contenton perods n the upln frame. Collsons are possble for such requests. In pollng based mode, BS assgns BW request opportuntes to the statons, and SSs reply by sendng BW requests. Other ways of contenton free BW request are BW stealng (sendng requests nstead of pacets) or pggybacng. When a UGS servce s setup, the BS s responsble for assgnng fxed sze perodc grants to the UGS flow; the SS s not allowed to explctly request BW for an already establshed UGS flow. UGS servce can be used for constant bt-rate (CBR) servce flows such as T1/E1. The ertps servce s smlar to UGS, except for ts use of dynamc and varable sze grants. In ths case, the BS provdes uncast grants of varable sze n an unsolcted manner le n UGS, thus savng the latency of a bandwdth request. The SS ndcates n pggybaced messages the requested sze for grants and the BS dynamcally adjusts the grant sze n the next cycle. ertps can be effcently used for VIOP flows wth slence suppresson. For the rtps and nrtps servces, the SSs can send BW requests n response to BW request opportuntes assgned by the BS. The dfference between these two modes s that the nrtps flows receve few pollng BW request opportuntes durng networ congeston and are allowed to use contenton requests, whle the rtps flows are polled regardless of the networ load, and frequently enough to meet the delay requrements of the servce flows. The rtps flows are not allowed to use contenton requests. The bandwdth assgnment to rtps flows s more flexble than UGS and can be used for VBR-le servce flows such as vdeo. The nrtps can be used for provdng better than best effort servces for applcatons such as bandwdth-ntensve fle transfer. Although ertps, rtps and nrtps flows requre throughput or btrate guarantee, ertps and rtps flows usually have addtonal requrements on delay. The delay performance of ertps can be better due to the fact that no explct BW request s needed. Best Effort (BE) flows are allowed to request bandwdth only through pggybacng and contenton mode. The parameters descrbng a traffc flow and ts QoS requrements n are dfferent for each class of traffc (except for ertps and rtps). Table I summarzes the mnmum requred nformaton for each class. Ths dfference between the type of QoS requrements ndcates the need for schedulng algorthms that are able to provde dfferent classes of qualty of servce, and not just dfferent levels of QoS or servce dfferentaton.

3 Table I. QOS PARAMETERS FOR DIFFERENT TRAFFIC CLASSES Traffc UGS ertps rtps nrtps BE Class QoS Parameter Maxmum Sustaned Traffc X X X X X Rate: R max Maxmum latency: D max X X Tolerated Jtter: J max X Mnmum Reserved Traffc Rate: R mn X (R max) X X X Traffc Prorty: X X Request/Transmsson Polcy X X X X X There are several physcal layer characterstcs of the standard that requre specal attenton when devsng a QoS soluton. The PHY operates n a multrate fashon. The multrate operaton s possble usng a user defned ln adaptaton mechansm that adjusts the modulaton and codng schemes, thus the transmsson rate, wth the objectve of mantanng certan relablty for the wreless ln (e.g., a BER of 10-5 ). For example, one set of rates provded by OFDM s , 21.7, 27.5, 41.5, 55.3, and 62.2 Mbps. Any of these transmsson rates can be used by each staton or the BS, and for any transmsson burst. The varaton n transmsson rate has a sgnfcant mpact on the farness and servce guarantee n the MAC, and has to be consdered n the scheduler desgn. III. ARCHITECTURE FOR FAIR SCHEDULING IN MAC Tradtonal schedulng solutons cannot be drectly employed n networs, due to specfc characterstcs of the MAC and PHY. These characterstcs nclude the dstrbuted multple access envronment wth the possblty of dynamc adjustment of upln/downln duratons, the multrate operaton of PHY, and the mult-class defnton of QoS requrements of connectons. We propose a detaled QoS scheme that taes nto account all these characterstcs and ams at provdng mult-class far QoS n a centralzed fashon. The promnent features of our soluton are the followng: (1) Combned Upln/Downln schedulng: our QoS framewor uses vrtual pacets and combnes the tas of schedulng upln and downln flows nto a central schedulng process that resdes n the BS; (2) Servce guarantee and schedulng dfferent classes of traffc: the proposed framewor uses a novel schedulng algorthm that allows long term servce farness whle reducng the delay of tme senstve ; and (3) Multrate operaton: the schedulng algorthm s desgned to provde ether temporal or throughput farness. To acheve the requred QoS, the MAC archtecture of the BS and SS should be specfed. We propose the archtecture depcted n Fg. 2 as a comprehensve QoS soluton. Ths archtecture requres a number of algorthms for ts components. To understand what components are needed, we consder the operaton of the MAC from the QoS perspectve. Snce the MAC s connecton orented, the frst step for transport s to setup a connecton. At connecton setup, a request s sent to the BS, declarng the traffc specfcatons of the stream that wll use ths connecton. The admsson control (AC) algorthm n the BS nspects the traffc specfcatons and admts or rejects the new connecton (stream) based on a user defned polcy. If the admtted flow s an upln flow, the traffc specfcaton s passed to a vrtual pacet generator (VPG) algorthm that generates vrtual pacets representng the actual upln pacets of the statons. The vrtual pacet generator also uses the upln queue length nformaton that s partly retreved through uncast poll generaton. Thus, a poll generaton (PG) algorthm s requred to schedule uncast request opportuntes (URO) for upln flows. Usng vrtual pacets, t s possble to use a sngle schedulng algorthm for schedulng both upln and downln flows. Ths scheduler can use any conventonal schedulng algorthm. We denote ths as the Inner BS Scheduler (IBS). The operaton of ths scheduler has a sgnfcant effect on the overall performance of the system. We propose a novel mult-class far queung scheme (MCFQ) for the nner scheduler. Actual or vrtual pacets that are served by the scheduler are passed to the frame bulder module whch constructs the MAC frame. Actual pacets are algned n a downln subframe, whle vrtual pacets are translated nto upln bandwdth grants called Transmsson Opportuntes (TO). UL and maps are then generated accordngly. In the SS sde, a smple schedulng algorthm that responds to bandwdth allocatons by the BS s enough. Ths algorthm sends pacets from the queues whch receve a BW allocaton from the BS, and f no such pacets exst, t can serve other queues usng a user defned prorty scheme. In fact, the QoS archtecture and algorthms at the BS are the major components requred for QoS n networs. The concepts and algorthms that are needed for enablng the BS archtecture are descrbed next. A. Combned Downln/Upln Schedulng, VPG Algorthm The schedulng dscplne must consder both upln and downln traffc for schedulng at all tmes. Downln pacets are avalable n the BS buffers and can be drectly scheduled, whereas upln pacets resde n the statons generatng these pacets and cannot be scheduled drectly. However, the BS can use traffc specfcatons, ndcated at connecton setup tme, to generate a local flow of vrtual pacets that represent the actual upln pacets. It s then possble to use a sngle scheduler to schedule these vrtual pacets, along wth downln pacets. The concept of vrtual pacets has been used before n other wors [14]. Vrtual pacets are generated usng the specfed patterns of upln traffc flow (e.g., specfed n dynamc servce addton messages at servce flow creaton and actvaton tme [1]) as well as the queue sze nformaton delvered through BW request messages. For example, for a voce call, a perodc flow of vrtual pacets smlar to the real traffc s generated. When there s slence n the voce call (detected usng pollng messages), vrtual pacet generaton s halted. When actvty s detected through BW request messages, vrtual pacet generaton resumes.

4 Base Staton (BS) DOWNSTREAM pacets Downln/Upln Servce Flow setup request (flow spec) Subscrber Staton (SS) UPSTREAM pacets Avalable tme for TX, T f Vrtual Pacet Generator (Algorthm VPG) Vrtual Pacets BS Inner Scheduler (algorthm IBS) Frame Bulder (Duplexng UL, Map) Flows Weghts Upln flow spec Admsson Control (algorthm AC) Flow Info Uncast Poll Generaton for BW Request (Algorthm PG) BW Request (Queue Len Info) Queue Length Info BW Request Generaton BW Req. Opp. BW Request. SS Scheduler MAC Upln Frame Access and Downln Frame Parsng PHY PHY Fg. 2. Archtecture of Far Servce Provsonng Mechansm for MAC The parameters used by VPG are taen from the mnmum set of parameters that should be provded for a connecton (see Table I). We mae a practcal assumpton that for audo and vdeo at least the followng set of parameters s provded: maxmum btrate (R max ), average btrate (R mn ), burst sze (b), and average pacet sze (p) or Servce Interval (T s ). We use these parameters, n addton to queue length nformaton, to generate vrtual pacets. The VPG algorthm reles on the pollng algorthm whch gathers queue sze nformaton (Q ) from statons. For as long as Q > 0, the VPG algorthm generates vrtual pacets of length p, at ntervals T s, to acheve the average rate R, where R max R R mn s determned by the admsson control scheme. If T s s avalable, pacet length can be calculated as R.T s, whereas f p s avalable, T s can be calculated as p/r. Wth each generated vrtual pacet, the projected queue length Q s reduced by the sze of the vrtual pacet. Q s ncreased by the amount of ncremental declared n each queue length update from the statons. Ths process ensures that vrtual pacets are only generated for the amount of avalable at the statons. Snce generaton rate may be varable n the staton, the algorthm mantans a budget parameter, g, to eep trac of the unused servce, and assgns t to the staton when t has more. Ths mechansm ensures upln flows receve a far share of the servce. The budget parameter s lmted to the allowed burst sze of the flow b n order to enforce a leay bucet shaped traffc. Ths algorthm s descrbed below Each loop runs n a separate thread. Loop QueueLenUpdate: If queue_length update s receved: (nc_q, ) Q = Q + nc_q End loop g = 0; // g s the credt for queue R = getrateac(); //AC determnes guaranteed rate: R max R R mn Loop VPG: g = g + R x T s ; //or +VP_length = p, wth T s = p/r ; f (g > b ) g = b ; f (Q > 0) { f (g > Q ) // BW grant should be lmted to Q VP_length = Q ; else // BW grant s lmted to the avalable budget g VP_length =g ; Generate_and_send_VP (, VP_length); g = g VP_length; Q = Q VP_length; // decrement to count for the // whch wll be sent n the upln subframe Wat T s ; }else Wat T s ; End loop; Snce upln pacet szes may vary and be dfferent from the vrtual pacet length, statons should fragment ther pacets (or zero pad) to ft them n the assgned TO. BE requests are drectly translated to vrtual pacets of the requested sze and no perodc VP generaton s needed.

5 B. Poll Generaton Algorthm: The uncast poll generaton scheme s a smple yet mportant part of the QoS archtecture, and s needed for rtps and nrtps classes. We propose an algorthm as follows: for rtps class, f a pggybaced uncast request for a queue has arrved durng the prevous frame, the BS does not send a URO to the correspondng staton. If no pggybaced request arrves, the BS generates UROs wth a perod equal to the servce nterval (SI) of the traffc. Snce rtps traffc (e.g., vdeo) s usually perodc, the servce nterval s nown. SI can also be calculated as a fracton of the maxmum latency. It must be noted that for rtps class, the standard permts the BS to ssue UROs even f prevous requests are stll unfulflled. For nrtps, traffc generaton s usually not perodc. For ths reason, the nterval between URO assgnments s mplementaton dependent. Also, f a prevous request of an nrtps flow s unfulflled, new UROs are not generated. The standard suggests the nterval between UROs (after the last BW assgnment ndcatng no at the staton) to be n the order of 1 second or less. When the networ s lghtly loaded, we reduce ths perod (T ) n order to reduce the response tme; whle t s ncreased to around 1 second when networ populaton grows. Therefore, for nrtps flows wth no unfulflled requests, T s calculated as follows: T =mn(1sec,θ*n nrtps ); where θ s a user defned parameter (e.g., 100msec) and N nrtps s the number of nrtps flows. C. Far Schedulng Algorthm The archtecture presented n ths paper allows for use of any conventonal schedulng algorthm as the central scheduler. For example, algorthms such as weghted round robn (WRR) or weghted far queung (WFQ) can be used [11]. Consderng that supports dfferent classes of QoS (.e., UGS, ertps, rtps and nrtps for upln, and smlar parameterzed flows for downln) wth dfferent requrements, the conventonal far schedulng algorthms cannot effcently provde delay and throughput guarantees and mantan farness at the same tme. To address these ssues we propose a new type of schedulng algorthm. QoS s usually specfed n terms of guaranteed bandwdth (throughput) and bounded delay. Ths s true for ertps and rtps class,.e., for vdeo and audo flows. However, for the nrtps class and non real tme applcatons, only guaranteed bandwdth s requred and large latency and jtter are tolerated. Farness n assgnng servce s also an mplct QoS requrement for flows whch are promsed the same level of servce by the networ. For example, users expect ther servce levels be mantaned whle the traffc of other users exceeds ts orgnally declared level. Bandwdth (throughput) guarantee s acheved through schedulng and admsson control. For the scheduler part, we need to ensure that each flow gets a weghted share of the channel capacty. Ths s usually acheved usng far algorthms such as WFQ. Usng these algorthms, the delay performance of each flow s strongly dependent on the weght (rate) assgned to t. On the other hand, nrtps flows do not have strct delay requrements, and the above algorthms treat both nrtps and rtps classes the same way. Ths s an undesrable property of the conventonal far algorthms. Better schedulng schemes can be devsed to use the fact that nrtps class does not requre delay guarantees. We brefly descrbe such an algorthm n ths secton. The proposed algorthm provdes short term farness amongst flows of the same class, and long term farness amongst dfferent classes of traffc. It also tres to reduce the delay bound for rtps flows, at the cost of ncreasng the delay for nrtps flows. D-UGS + Vrtual UGS/ertPS mfq Start/Fnsh Tme Stampng xx: any scheduler mfq: modfed SFQ MCFQ: Mult-Class Far Queueng D-rtPS + vrtual rtps MCFQ Elgblty Tme Stamp MCFQ Start/Fnsh TmeStampng Prorty Schedulng (UGS,ertPS > (n)rtps > BE) D-nrtPS + vrtual nrtps D-BE + vrtual BE mfq MCFQ xx Fg. 3. Inner Scheduler Archtecture Snce downln connectons can use any traffc descrpton, ncludng the descrptons specfed for the fve upln servces, we can classfy downln flows nto the same fve servce types. However, ertps class uses the same descrptors as rtps (Table I) and s not needed for the downln drecton. Therefore, we assume 4 servce types for the downln drecton, dentfed as: D-UGS, D-rtPS, D-nrtPS and D-BE. We note that n the centralzed scheduler, upln and downln servces of the same type are grouped together; therefore we drop the D- prefx where there s no dstncton. Upln ertps class s grouped wth UGS due to ts requrement for unsolcted UTO grant. The archtecture of the proposed scheduler s presented n Fg. 3. Ths desgn uses 2 layers of schedulng. Three far schedulng algorthms are used for schedulng flows of the same type for classes of combned upln UGS and ertps (whch have unsolcted UTO grants) and downln UGS le flows, combned downln or upln rtps and nrtps flows, and downln or upln BE. A prorty scheduler s used n the second layer to serve the pacet that s selected by the far schedulng algorthms. The prorty scheduler gves the hghest prorty to UGS/ertPS/D-UGS (by defnton, generated UGS/ertPS vrtual pacets should be mmedately served). Combned downln or upln rtps/nrtps s gven hgher prorty than BE. We emphasze that farness s provded amongst flows of the same class (e.g., amongst combned downln or upln rtps/nrtps flows, or amongst combned D- UGS/UGS/ertPS). The BE traffc s only served when no traffc of other QoS classes exsts. Avodng starvaton for BE requres AC and

6 VPG modules lmtng the amount of both upln traffc assgned to the QoS classes. Upln traffc shapng s smply acheved by the VPG generatng conformng vrtual flows; downln traffc shapng requres a traffc shaper n the BS. Snce no QoS guarantee s needed wthn the BE class, any scheduler can be used amongst BE flows; we use WRR to provde long term farness amongst BE flows. UGS flows are scheduled usng a far queung algorthm. We propose to use a modfed Start-tme Far Queung (SFQ) [13]. The modfcaton we propose s to change the SFQ algorthm to acheve ether temporal or throughput farness, when a multrate PHY s consdered. Ths modfcaton s dscussed later n ths secton. Snce rtps and nrtps classes requre BW guarantee, we combne these two classes and present a novel mult-class far queueng algorthm (MCFQ) to provde throughput or BW guarantee for both classes, whle provdng better delay performance for the rtps (both upln and downln) class. We note that most QoS demandng applcatons wll use these two classes. The MCFQ algorthm s descrbed below. MCFQ Algorthm MCFQ s based on the archtecture shown n Fg. 3. In MCFQ, smlar to SFQ, pacets are tagged wth start and fnsh tme tags, denoted here as (S, F ) for pacet of flow, as follows: 1 S = max( F, V ( t)) (1) F = S + L / ϕ where, V(t), s a vrtual tme functon that tracs the overall progress of the system, φ are weghts assgned to each flow, and L s the length of the pacet. Addtonally, rtps pacets are tme stamped wth an elgblty tme, whch s desgned to allow a burst of sze b to be served for rtps flows ahead of the usual SFQ order, as s explaned later. The elgblty tme s defned as: E = S b ϕ (2) The elgblty s determned by comparng ths value to the vrtual tme of the system. Thus a pacet s deemed elgble at tme t f E V (t), and nelgble otherwse. Vrtual tme s mantaned to trac the progress of the system, and can be used to smulate a generalzed processor sharng (GPS) flud server (smlarly done n WFQ or SFQ). MCFQ adjusts vrtual tme dfferently from SFQ. In MCFQ, n each round of servce, the scheduler serves the elgble rtps pacet wth the smallest start tme; f no elgble rtps pacet s found, any pacet wth the smallest start tme tag s served. The vrtual tme s set to the start-tme tag of the head of lne pacet wth the smallest start tme: V ( t) = mn{ S : HoL pacet of queue j & j all queues} (3) j Note that ths pacet mght be an nrtps pacet and not actually n transmsson yet; ths step maes the algorthm behavor very dfferent from other GPS based schemes. Ths choce of vrtual tme smulates the behavor of SFQ n the long run, although pacets are not served n SFQ order n the short term. Ths operaton ensures that the scheduler s far n the long term and does guarantee btrate, whle at the same tme the rtps flows get much better delay performance compared to when a sngle class scheduler (e.g., SFQ) s used. Settng the E accordng to (2) has the effect that a burst of rtps, less than b n length and arrvng at tme t, wll be served before all nrtps flows watng at ths tme; the reason s that the pacets n the burst wll have E = S b ϕ < V (t) and are elgble before all other nrtps pacets untl V(t) ncreases wth the servng of the next nrtps or nelgble rtps flow wth hgher start tme value. As a result, the pacets n the burst wll only have to compete wth other elgble rtps flows. Pacets that arrve mmedately followng ths burst, wll not be elgble, snce ther start tme stamp wll be ahead of V(t) wth more than b /φ, mang them nelgble for servce wth the current burst of the same flow. The effect s a natural shapng of the flow whch wll result n long term farness amongst rtps and nrtps flows. A rgorous study of the propertes of the MCFQ algorthm follows smlar arguments as n [13][15] and wll be presented n our upcomng publcatons. D. Multrate operaton of the PHY and Farness of Schedulng The multrate operaton of the PHY sgnfcantly affects the farness of the servce provded to flows [10][15]. Pacets that are transmtted usng low transmsson rates tae longer tme on the channel, and lower the overall capacty of the system, negatvely affectng other flows. Ths stuaton may be deemed unfar f we defne farness n terms of the servce tme. Ths type of farness s also called temporal farness for whch the goal s to provde the statons (flows) wth the same (weghted) amount of servce tme regardless of ther PHY transmsson rate. In contrast to temporal farness, we can also defne farness n terms of the throughput that s assgned to statons (flows) n a gven duraton of tme (usually longterm); ths s called throughput farness. In ths case, farness s acheved f all statons acheve the same (weghted) throughput n a gven duraton of tme [15]. The schedulng process can be desgned to provde ether temporal or throughput farness, based on the overall performance objectve. A detaled analyss of algorthms for ths purpose, are out of the scope of ths paper and can be found n [15]; however for the purpose of provdng a complete pcture for the schedulng soluton, we present the modfcatons requred for farness provson. To provde an nsght, we tae the SFQ example and then extend t to MCFQ. For throughput farness, one can smply use the SFQ algorthm. Ths algorthm s nherently throughput-far. For temporal farness, SFQ can be modfed to Start tme Temporal Far Queung (STFQ) algorthm, by modfyng the fnsh tme from (1) to the followng: F = S + L /( ϕ C ) (4) where C s the PHY transmsson rate at whch pacet of flow s transmtted. The MCFQ algorthm presented n the prevous secton s by desgn throughput far. To convert t to a temporal far

7 algorthm, we convert the bt-based servce tracng parameters of the algorthm to tme-based components. For example, smlar to STFQ, the fnsh tme equaton changes to (4), and the elgblty tme equaton changes to the followng: E = S b ( ϕ. C ) (5) These changes are enough for convertng the algorthm to a temporal far algorthm, snce the progress of each queue, n terms of the receved servce, s only traced n the fnsh tme and elgblty tme equatons. Although multrate operaton provdes relable lns most of the tme, occasonal pacet loss s stll possble. The lost pacets are retransmtted n the MAC. In the case of throughput far schedulng, the pacet transmsson s smply retred, and the algorthm remans throughput far. For temporal far case, the tme stamps of the retransmtted pacet s recalculated as f the pacet has just arrved to the queue head of lne. Ths readjustment s performed easly for SFQ, snce tme stampng pacets can be done for head of lne pacets only [14]. E. Admsson Control Any admsson control algorthm may be used n the archtecture n Fg. 2, wthout need for modfyng archtectural desgn. For example, algorthms based on btrate reservaton may be used. Such algorthms reserve resources to guarantee the mnmum btrate (equal to the average btrate of the flow), the pea rate (maxmum sustaned rate), or a value n between, determned based on stochastc guarantees. These algorthms have already been extensvely dscussed n research lterature and are not repeated here [9]. However, we pont out that the weght assgnment for flows s done n the admsson control module. Moreover, the admsson control module should be aware of the decson to use temporal or throughput farness, and act accordngly. IV. PERFORMANCE EVALUATION To demonstrate the effectveness of the proposed solutons we conducted a seres of smulaton experments. These experments hghlght the abltes of the proposed soluton n provdng the followng servces: 1) far servce guarantee for upln/downln flows, 2) far guaranteed servces n multrate networs, 3) provdng preferental delay performance for real tme (e.g, rtps) flows, whle guaranteeng far servce for nrtps as well. We mplemented a detaled dscrete event smulator n C for the MAC n TDD mode, and observed the delay and throughput performance. We assumed a multrate PHY ( OFDM). In the frst experment we setup a networ wth 10 upln and 10 downln flows (all of rtps type). Pacets of unformly dstrbuted sze n [500,1500] were generated wth exponental nter-arrval tme of 5.3 msec (Posson arrval process). The number of bts served for each flow and the delay performance of the flows are depcted n Fg. 4. The slope of the curve for the number of served bts gves the throughput. It s seen that both type of flows acheve the same throughput. It s also seen that the upln flows ncur more delay, whch s expected snce upln pacets are only served after a report of ther arrval s receved by the BS. These performance curves demonstrate that the scheme based on vrtual pacets s ndeed able to guarantee btrate, despte the fact that vrtual pacets are generated almost perodcally and have a sze dfferent from upln pacets. Fg. 4. Throughput and Delay performance, upln vs. downln In the second experment we evaluated the throughput n a multrate envronment for the same flows as n the prevous example. Fg. 5 shows the served bts curve for both temporal far (STFQ) and throughput far (SFQ) algorthms. STA1, STA2 and STA3 were transmttng at PHY speeds of 64, 32, and 21.6 Mbps, respectvely. Fg. 5. Throughput Farness vs. Temporal Farness (PHY rates of Statons 1,2 and 3: 65, 32.5, 24.4 Mbps) It s seen n Fg. 5, that usng a temporal far scheduler, each flow gets a throughput proportonal to ts PHY btrate. Wth a throughput far algorthm, all flows get the same throughput, whch s less than the maxmum throughput a staton would get n the temporal far case. Ths reducton n throughput s due to the fact that the reduced capacty s now shared amongst all statons, and not just the statons wth lower transmsson rates.

8 throughput). Pacet szes for nrtps flows were set as n the prevous example, and a 20KB burst sze was used for rtps flows n MCFQ. In conventonal far schedulng, all flows are assgned weghts n a way that they all receve a far guaranteed throughput. The results are demonstrated n Fg. 6. Snce the vdeo flow has a hgher burst sze than the other flows, t ncurs a larger delay when conventonal far schedulng algorthms le SFQ are used; ths s despte the fact that all flows receved ther requested share of throughput or BW (shown n Fg. 7) n long term. When networ s not heavly loaded (50% n Fg. 6) all flows ncur small delay. When networ gets close to ts full capacty (75% and above), MCFQ demonstrates a sgnfcant better delay performance for rtps flows, whle stll mantanng long term farness and throughput guarantee for all. The better delay performance for rtps flow, n MCFQ, comes at the cost of delay ncrease for nrtps flows, whch s perfectly acceptable snce nrtps flows only requre btrate or BW guarantee. The bottom plot n Fg. 7 shows how rtps flow n MCFQ receves better servce (jump n served bts curve at 1 sec ntervals where I-frames are sent), but has to gve up the extra servce n long run (lower slope and the resultng sawtooth effect on curve). In Fg. 8 and Fg. 9 we plotted the delay nstances of each pacet for the 100% load scenaro, to see n detal how each pacet s ncurrng delay and that the flows queues are stable (the other nrtps queues wll not be stable snce ther load s selected to saturate the networ to 100%). Fg. 6, Fg. 8 and Fg. 9 all show that the maxmum ncurred delay for rtps flows s greatly reduced f MCFQ s used. Fg. 6 Delay performance for dfferent load levels, MCFQ vs. SFQ, from top (50%, 75%,95%,100%) In the last experment, we observed the delay performance of the mult-class far scheduler proposed n ths paper. We compared the results wth those obtaned by other far algorthms such as SFQ. For ths experment we consdered a foreman H.264 vdeo flow (2Mbps) beng served as an rtps traffc n a networ loaded wth 19 other nrtps bacground traffc. We set one nrtps flow to 2Mbps to compare t wth vdeo; the other flows had a varable load n order to observe the effect of dfferent levels of load on MCFQ and SFQ delay performance. All flows had equal weght (or reserved Fg. 7 Served bts for rtps and nrtps flows, usng MCFQ and SFQ algorthms; top plot (smulaton tme vew), bottom plot (zoomed n)

9 REFERENCES Fg. 8 Pacet delay for rtps vdeo flow, MCFQ vs. SFQ; (bacground load 100%) Fg. 9 MCFQ delay nstances for one smulaton run, nrtps flows vs. rtps (vdeo) flow; (load 82%) V. CONCLUSION Although the standard provdes features for QoS provsonng, t ntentonally does not mandate or provde a complete QoS soluton. Such a soluton requres many mechansms and algorthms that are left to be developed by vendors. In ths paper, we presented a comprehensve archtecture for QoS provsonng n networs. We dentfed several mechansms n ths archtecture that must be mplemented before a fnal QoS soluton s avalable. The man contrbutons of ths paper are the algorthms that are developed as enablng components of the presented archtecture. The mult-class far schedulng algorthm that has been presented here s able to support dfferent classes, as well as, levels of QoS. We have also presented measures to adjust the schedulng algorthm for temporal or throughput farness provson n multrate networs. The archtectures and algorthms presented n ths paper provde a good example for constructng a QoS enabled WMAX access networ. Nevertheless, these algorthms can also be used n other access networs wth smlar structural requrements (upln and downln dfference, multrate PHY or dfferent classes of traffc). [1] IEEE Standard for Local and Metropoltan Area Networs Part 16: Ar Interface for Fxed Broadband Wreless Access Systems. ANSI/IEEE Std (Revson of IEEE Std ). [2] IEEE Standard for Local and Metropoltan Area Networs Part 16: Ar Interface for Fxed Broadband Wreless Access Systems. Amendment 2: Physcal and Medum Access Control Layers for Combned Fxed and Moble Operaton n Lcensed Bands. ANSI/IEEE Std e-2005 [3] Wreless LAN Medum Access Control (MAC) and Physcal Layer (PHY) specfcatons. ANSI/IEEE Std : 1999 (E) Part 11, ISO/IEC , [4] M. Hawa and D. W. Petr, Qualty of Servce schedulng n cable and broadband wreless accesss systems, Tenth IEEE Internatonal Worshop on Qualty of Servce, May 2002 Page(s): [5] C. Ccconett, A. Erta, L. Lenzn, and E. Mngozz, Performance Evaluaton of the IEEE MAC for QoS Support, IEEE Transactons on Moble Computng, Vol. 6, No. 1, January [6] D.-H. Cho, J.-H. Song, M.-S. Km, and K.-J. Han, Performance Analyss of the IEEE Wreless Metropoltan Area Networ, Proc. Frst Int'l Conf. Dstrbuted Framewors for Multmeda Applcatons (DFMA '05), pp , Feb [7] O. Gusa, N. Olver, and K. Sohraby, Performance Evaluaton of the Medum Access Control Layer, Lecture Notes on Computer Scence, vol. 3280, pp , [8] IEEE Standard e/ Amendment 8, Medum Access Control (MAC) Qualty of Servce (QoS) Enhancements, July 2005 [9] M. Ghader and R. Boutaba, Call admsson control for voce/ ntegraton n broadband wreless networs, IEEE Transactons on Moble Computng, Vol. 5, No. 3, March [10] Y. Yuan, D. Gu, W. Arbaugh, J. Zhang, Achevng Far Schedulng Over Error-Prone Channels n Multrate WLANs,Wreless Networs, Communcatons and Moble Computng, 2005 Internatonal Conference on Volume 1, June 2005 pp [11] A. Pareh, R. Gallager, A Generalzed Processor Sharng Approach to Flow Control n Integrated Servces Networs: The Sngle-Node Case," IEEE/ACM Trans. on Networng, vol.1, no. 3, pp , June 93. [12] J.C.R. Bennett,Z. Hu, WF 2 Q: Worst-Case Far Weghted Far Queueng, INFOCOM '96. Ffteenth Annual Jont Conference of the IEEE Computer Socetes. Networng the Next Generaton. Proceedngs IEEE [13] P. Goyal; H.M. Vn; C. Hachen, Start-Tme Far Queueng: A Schedulng Algorthm For Integrated Servces Pacet Swtchng Networs, Networng, IEEE/ACM Trans. on, Vol 5, Issue 5, Oct pp [14] Y. Pourmohammad Fallah, H. Alnuwer "Hybrd Pollng and Contenton Access Schedulng n IEEE e WLANs, Journal of Parallel and Dst. Comp., Elsever, Vol 67, Issue 2, Feb. 2007, pp [15] Y. Pourmohammad Fallah, H. Alnuwer, Analyss of Temporal and Throughput Far Schedulng n Mult-Rate IEEE e WLANs, Computer Networs, Elsever, Vol. 52, Issue 16, November 2008,pp