An Effcent Schedulng For Dverse QoS Requrements n WMAX by Xaojng Meng A thess presented to the Unversty of Waterloo n fulfllment of the thess requrement for the degree of Master of Appled Scence n Electrcal and Computer Engneerng Waterloo, Ontaro, Canada, 2007 Xaojng Meng, 2007
AUTHOR'S DECLARATION I hereby declare that I am the sole author of ths thess. Ths s a true copy of the thess, ncludng any requred fnal revsons, as accepted by my examners. I understand that my thess may be made electroncally avalable to the publc.
Abstract WMAX s one of the most mportant broadband wreless technologes and s antcpated to be a vable alternatve to tradtonal wred broadband technques due to ts cost effcency. Beng an emergng technology, WMAX supports multmeda applcatons such as voce over IP (VoIP), voce conference and onlne gamng. It s necessary to provde Qualty of Servce (QoS) guaranteed wth dfferent characterstcs, qute challengng, however, for Broadband Wreless Access (BWA) networks. Therefore, an effectve schedulng s crtcal for the WMAX system. Many traffc schedulng algorthms are avalable for wreless networks, e.g. Round Robn, Proportonal Farness (PF) scheme and Integrated Cross-layer scheme (ICL). Among these conventonal schemes, some cannot dfferentate servces, whle some can fulfll the servce dfferentaton wth a hgh-complexty mplementaton. Ths thess proposes a novel schedulng algorthm for Orthogonal Frequency Dvson Multplex/Tme Dvson Multple Access (OFDM/TDMA)-based systems, whch extends the PF scheme to multple servce types wth dverse QoS requrements. The desgn objectve s to provde dfferentated servces accordng to ther QoS requrements, whle the objectve can be acheved by adjustng only one unque parameter, the tme wndow for evaluatng the average throughput. By extensve smulaton, t s shown that the proposed schedulng algorthm explots the advantage of the PF scheme, enhancng the throughput, and dstngushes the servces n terms of the average delay. Afterward, we prove the superorty of the new scheme over the conventonal ones by showng smulaton results.
Acknowledgements Frst and foremost, I would lke to express my grattude to my supervsors, Professor Pn- Han Ho and Kanam Thomas Wong, for ther gudance and support durng my gradate studes. Wthout ther help, the achevements n my research would never have been possble. I would also lke to thank Professor Murat Uysal and Professor Langlang Xe for beng the readers of ths thess and for ther nsghtful comments and suggeston. Many thanks to my frends and all members n my research group for ther frendshp, support and helpful dscusson. Thanks to the admnstratve support staff, namely Lsa Hendel, Wendy Boles and Karen Schooley. I am forever ndebted to my parents for ther constant support and encouragement throughout the past years. Ths thess s devoted to them. v
Table of Contents Chapter Introducton.... Motvaton....2 Research Challenges...2.3 Man Contrbutons...3.4 Thess Outlne...3 Chapter 2 Wreless Metropoltan Area Networks Overvew... 4 2. PHY Technology n IEEE 802.6...5 2.. Broadband Wreless Access Background...6 2..2 Adaptve Modulaton and Codng (AMC)...8 2.2 MAC Layer n IEEE 802.6...9 2.2. MAC Support of PHY... 2.2.2 Schedulng Servce...2 2.2.3 Multuser Dversty...3 Chapter 3 An Overvew of Schedulng Algorthms... 5 3. Round Robn (RR)...5 3.2 Proportonal Farness (PF)...6 3.3 Integrated Cross-layer Schedulng...6 Chapter 4 Scheduler Desgn... 9 4. System Model...9 4.2 Channel Fadng Model...20 4.3 Proposed Algorthm Schedulng...2 4.3. Prorty Functon...2 4.3.2 Tme Slot Allocaton...23 Chapter 5 Analyss and Smulaton Results... 24 5. Performance Analyss...24 5.2 Smulaton Results...26 5.2. System Parameter Settng and Assumpton...26 5.2.2 Smulaton Results...27 v
5.2.3 Comparson wth Other Schedulng Schemes...39 Chapter 6 Concluson and Future Work... 44 6. Concluson...44 6.2 Extensblty...44 6.3 Future Work...45 Appendx Lst of Abbrevatons... 46 Bblography... 49 v
Lst of Fgures 2.. Fg. 2. Fxed PMP archtecture network... 6 2.. Fg. 2.2 OFDM/TDMA & OFDMA... 8 2.2. Fg. 2.3 IEEE 802.6 protocol layerng... 0 2.2. Fg. 2.4 TDD frame structure n IEEE 802.6... 2.2. Fg. 2.5 TDD downlnk subframe structure n IEEE 802.6... 2 2.2.3 Fg. 2.6 Packet Scheduler Structure... 4 5.2.2. Fg. 5. Comparson for Prorty value under dfferent T sets... 28 5.2.2.2 Fg. 5.2 Comparnson for Average delay under dfferent T sets (a)-(g)... 30-32 5.2.2.2 Fg. 5.3 Comparson for Throughput under dfferent T sets (a)-(d)... 33 5.2.2.4 Fg. 5.4 Comparnson for Average delay under dfferent T sets... 35-36 5.2.2.4 Fg. 5.5 Comparnson for Throughput under dfferent T sets... 37-38 5.3.2 Fg. 5.6 System performance under RR and PF... 4 5.3.2 Fg. 5.7 System performance under ICL... 43 v
Lst of Tables 2.. Table Ar nterface nomenclature... 7 2..2 Table 2 Transmsson Modes n IEEE 802.6... 9 5.2.. Table 3 Dstance from SS to the BS... 26 5.2..2 Table 4 Input servce flow of each SS... 27 5.2.2.4 Table 5 T Settng and the Analyss Values... 3 5.2.3 Table 6 Parameter Settng n ICL... 42 v
Chapter Introducton. Motvaton WMAX (Worldwde nteroperablty for Mcrowave Access) s one of the most emergng technologes for Broadband Wreless Access (BWA) n metropoltan areas by provdng an exctng addton to the current broadband technques for the last-mle access. It s demonstrated that WMAX s a vable alternatve to the cable modem and DSL technologes due to ts hgh resource utlzaton, easy mplementaton and low cost. Furthermore, WMAX not only enhances the exstng features of the compettve cabled access networks, but provdes hgh data rate applcatons wth a varety of Qualty of Servce (QoS) requrements. We are reachng the goal of realzng a unque wreless network to cover a bg area. In a large scale wreless network, the rado resource must be shared among multple users. The bandwdth allocaton algorthms have been desgned for the effcent utlzaton of the scarce rado resource. In addton, to support multmeda traffcs, the Medum Access Control (MAC) protocols wll co-ordnate the transmsson of traffc flows. The channel characterstcs of users and traffc flow requrements are largely dverse, motvatng us to desgn an effcent MAC layer protocols that can mprove the system performance due to the channel and traffc dynamcs.
.2 Research Challenges Schedulng algorthms serve as an mportant component n any communcaton network to satsfy the QoS requrements. The desgn s especally challenged by the lmted capacty and dynamc channel status that are nherent n wreless communcaton systems. To desgn an MAC layer protocol whch can optmze the system performance, the followng features and crtera should be concerned. [-3] Bandwdth utlzaton Effcent bandwdth utlzaton s the most mportant n the algorthm desgn. The algorthm must utlze the channel effcently. Ths mples that the scheduler should not assgn a transmsson slot to a connecton wth a currently bad lnk. QoS requrements The proposed algorthm should support dfferent applcatons to explot better QoS. To support delay-senstve applcatons, the algorthm provdes the delay bound provsonng. The long-term throughput should be guaranteed for all connectons when the suffcent bandwdth s provded. Farness The algorthm should assgn avalable resource farly across connectons. The farness should be provded for both short term and long term. Implementaton complexty In a hgh-speed network, the schedulng decson makng process must be completed very rapdly, and the reconfguraton process n response to any network state varaton. Therefore, the amount of tme avalable to the scheduler s lmted. A low-complexty algorthm s necessary. Scalablty The algorthm should operate effcently as the number of connectons or users sharng the channel ncreases. Our protocol desgn s desrable to fulfll all of the above features. 2
.3 Man Contrbutons The major objectve of ths thess s to develop an effectve yet smple QoS schedulng scheme for mult-servce networks that can be deployed and mplemented wth the less overhead. We explore the lmtaton on the proportonal farness (PF) schedulng scheme, and propose to extend the PF scheme for delay dfferentaton n IEEE 802.6 networks (or referred to as WMAX). For meetng dfferent delay constrants on each servce type, T can be manpulated to serve ths purpose. The analyss results are gven by the smulatons, whch have shown that the proposed algorthm can dfferentate servces n terms of the delay constrants over a large scale network..4 Thess Outlne Chapter 2 ntroduces the background of IEEE802.6 Physcal Layer (PHY) and Medum Access Control (MAC) layer protocols. An overvew of prevous related work s presented n Chapter 3. Chapter 4 descrbes the developed protocol. The analytcal and smulaton results by usng MATLAB are valdated n Chapter 5. Fnally, the conclusons are drawn n Chapter 6. 3
Chapter 2 Wreless Metropoltan Area Networks Overvew In the early 2000 s, BWA n Metropoltan Areas has been recognzed as one of the most promsng technologes that wll be wdely deployed n the world. In order to rapdly converge on a worldwde standard, several standards have been publshed. A number of optons are provded n the IEEE 802.6 famly. [5-6] IEEE 802.6a: The standard specfes the operaton from 2GHz to GHz, both lcensed and lcense exempts. Because the sgnals at lower frequency can penetrate barrers and thus a lne-of-sght connecton between the transcever and recever s not requred, most commercal nterests have focused manly on the lower frequency ranges. Under ths premse, IEEE 802.6a standard was thus completed n January 200. It enables the WMAX mplementatons wth better flexblty whle mantanng the data rate and transmsson range. IEEE 802.6a also supports mesh deployment, whch can extend the network coverage and ncrease the overall throughput. IEEE 802.6b: Ths extenson ncreases the spectrum to the 5 and 6 GHz frequency bands, whch provdes QoS guarantee to ensure prorty transmsson for real-tme applcatons and to dfferentate servce classes for dfferent traffc types. 4
IEEE 802.6c: As the Work Group s ntal nterest, IEEE 802.6c defnes a 0 to 66 GHz system profle that standardzes more detals of the technology. These hgh frequency bands have more avalable bandwdth, but the sgnals cannot dffract the obstacles and requre lne of sght deployment. IEEE 802.6d: Approved n June 2004, IEEE 802.6d upgrades the 802.6a standard. Ths extenson ams to mprove performance for 802.6 especally n the uplnk traffc. IEEE 802.6e: Ths technology standardzes networkng between fxed base statons (BSs) and moble base statons (MSs), rather than just between base statons and fxed recpents. IEEE 802.6e enables the hgh-speed sgnal handoffs necessary for communcatons wth users movng n vehcles. It promses to support moblty up to speeds of 70-80m/h. The subscrber statons (SSs) could be personal communcaton devces such as moble phones and laptops. We only concentrate on some basc characterstcs of IEEE 802.6d PHY and MAC protocols that are necessary for downlnk schedulng algorthm desgn n the fxed network archtecture. In the followng sectons, an overvew on IEEE 802.6 PHY subsystems s provded. 2. PHY Technology n IEEE 802.6 IEEE 802.6 s a unversal standard comprehendng varous types of network archtecture. IEEE 802.6 defnes two dfferent network topologes each wth a specfc MAC protocol: the pont to multpont (PMP) mode and mesh mode. The mesh mode s optonal n IEEE 802.6e, where data can be routed drectly between two SSs. In the PMP mode, a central BS s capable of handlng multple ndependent SSs smultaneously. It does not need to coordnate wth other statons. Nowadays, most WMAX systems are equpped wth the PMP mode where traffc only occurs between a BS and ts SSs [7]. 5
BS Fg. 2. Fxed PMP archtecture network 2.. Broadband Wreless Access Background Several frequency bands for the ntal 802.6 products have been dentfed. In IEEE 802.6a-200, the frequency s addressed from 0 to 66 GHZ, whch s avalable all over the world. Due to hgher frequency, Lne-of-Sght (LOS) propagaton s a necessty. For a resdental applcaton, roof tops may be too low for a clear sght lne to a BS. We must consder the multpath propagaton affecton. Recently, more nterest s n the 2-GHz projector. Desgn of the 2- GHz PHY s drven by the need for non-los (NLOS) operatons. The standard defnes three dfferent ar nterfaces that can be used to provde a relable end-to-end lnk: SCa: A sngle-carrer modulated ar nterface. OFDM: A 256-carrer orthogonal-frequency dvson multplexng (OFDM). Multple access of dfferent SSs s tme-dvson multple access (TDMA)-based. OFDMA: A 2048-carrer OFDM scheme. But a subset of the carrers can be assgned to an ndvdual user. It s referred to be OFD multple accesses. 6
Table summarzes the nomenclature for the varous ar nterface specfcatons n standard. Table Ar nterface nomenclature Desgnaton Applcablty Duplexng alternatve WrelessMAN-SC 0-66GHZ TDD&FDD WrelessMAN-SCa Below GHZ,lcensed band TDD&FDD WrelessMAN-OFDM Below GHZ,lcensed band TDD&FDD WrelessMAN-OFDMA Below GHZ,lcensed band TDD&FDD Among these three ar nterfaces, the two OFDM-based systems are more sutable for NLOS due to the smplcty of the equalzaton process for multcarrer sgnals [8]. In a multple access communcaton system, transmsson resources are shared among multple users such that a resource management scheme s requred. TDMA and Frequency Dvson Multple Access (FDMA) are two well-known technques for resource management based on the prncple of tme sharng and frequency sharng, respectvely. When combned wth OFDM (Orthogonal Frequency Dvson Multplexng), they are called OFDM-TDMA and OFDMA (OFDM Access), respectvely. The tme and subcarrer assgnment s llustrated n Fg. 2.2. All profles currently defned by the WMAX Forum specfy the 256-carrer OFDM PHY. For ths reason, the study focuses prmarly on the 256-carrer OFDM/TDMA ar nterface, where each SS can take an OFDM symbol wth all the subcarrers wthn a tme slot exclusvely. The access by multple users s realzed along the tme doman. The advantage of OFDM/TDMA s that the number of physcal tme slots and the number of codes assgned are adjustable. It leads to dfferent data rate armed wth adaptve modulaton and codng (AMC) avalable n PHY. 7
Fg. 2.2 (a) OFDM/TDMA Fg. 2.2 (b) OFDMA 2..2 Adaptve Modulaton and Codng (AMC) The man objectve of adaptve modulaton and codng s to compensate for rado channel nstablty. It has been shown n [9-2] that adaptve modulaton can effectvely mprove the bt error rate (BER) performance on rado channels whch had suffered from shadowng and fadng. The modulaton schemes defned n the IEEE 802.6 standard n the downlnk and uplnk are bnary phase shft keyng (BPSK), quaternary PSK (QPSK), 6-Quarter Ampltude Modulaton (QAM), and 64-QAM. The system could yeld the best performance f the swtchng thresholds are selected carefully. Gven the BER less 6 than0, the algorthm of choosng the optmal transmsson mode s n accordance wth Table 2 [22]. For WMAX, the selecton of a modulaton scheme bascally takes advantage of the rado channel measurements extracted by the SS whch exercses the Channel State Informaton (CSI) and the retransmsson procedure. The measurement on the Sgnal to Nose Raton (SNR) s performed on each frame preamble. 8
Table 2 Transmsson Modes n IEEE 802.6 Modulaton Codng rate W bts/symbol Recever SNR (db) BPSK /2 0.5 6.4 /2.0 9.4 QPSK 3/4.5.2 /2 2.0 6.4 6 QAM 3/4 3.0 8.2 2/3 4.0 22.7 64 QAM 3/4 4.5 24.4 Based on a perfect channel measurement whch can match the transmtter parameters to tme varyng channel condtons, a proper modulaton and codng method can be chosen for the upcomng transmsson so that the user bt rate can be maxmzed. AMC has been used to provde hgh-speed data transmsson by many standard wreless networks, such as IEEE 802./6 and 3GPP/3GPP2 [23]. 2.2 MAC Layer n IEEE 802.6 The MAC layer of IEEE 802.6 s desgned to serve sparsely dstrbuted statons wth hgh data rates, where the SSs are not requred to lsten to the other statons lke the MAC n IEEE 802.. The BS schedules the transmssons of the correspondng SSs n advance. The MAC of WMAX s reservaton-based and contenton-free. The SSs need to contend only when they access the channel for the frst tme at the connecton admsson control stage. The reservaton-based resource allocaton allows the WMAX BS to serve a large number of SSs as well as the guarantee of QoS n the connecton level for both uplnk and downlnk traffc. Compared wth 802.6, Wreless Local Area Networks (WLAN) based on IEEE 802. termnals usually have ntermttent traffc that contends every tme before transmttng, where the effcency s sgnfcantly mpared when more statons enter the network. In such a contenton based resource reservaton scheme, QoS could 9
hardly be consdered n the early standard untl the advent of 802.e. However, most WLAN networks deployed nowadays do not employ any QoS mechansm [24]. The man purpose of the MAC protocol s the sharng of rado channel resources among multple accesses of dfferent users. In IEEE 802.6, the MAC layer s dvded nto three sublayers: the servce-specfc convergence, common part sublayer, and securty sublayer. The prmary task of servce-specfc convergence sublayer s to classfy external servce data unts (SDU) and assocate each of them wth a proper MAC servce flow dentfer and connecton dentfer. The MAC layer protocol s flexble and effcent over dfferent traffc types. The common part sublayer s ndependent of the transport mechansm, whch s the kernel bearng all the MAC characterstcs. It s responsble for fragmentaton and segmentaton of each MAC SDU nto MAC protocol data unts (PDUs), system access, bandwdth allocaton, connecton mantenance, QoS control, and schedulng transmsson, etc. The MAC also contans a separate securty sublayer handlng authentcaton, secure key exchange, and encrypton. Servce specfc Convergence Sublayer MAC PHY MAC Common Part Sublayer Securty Sublayer Physcal Layer Fg. 2.3 IEEE 802.6 protocol layerng 0
2.2. MAC Support of PHY The basc dstncton of MAC protocol s the duplexng technques of uplnk and downlnk. The choce of duplexng technques may affect PHY parameters as well as mpact the features that can be supported. There are two approaches to mplement t. ) Frequency Dvson Duplex (FDD): In an FDD system, the uplnk and downlnk channels use separate subcarrers, whch allows the termnals to transmt and receve smultaneously. The adopton of fxed duraton frames n both uplnk and downlnk smplfes the desgn of bandwdth allocaton algorthms. 2) Tme Dvson Duplex (TDD): In ths paper, TDD framng s llustrated n Fg. 2.4. The uplnk and downlnk transmssons share the same frequency whle beng allocated n each TDD frame accordng to an adaptve threshold. One TDD frame (Fg. 2.4) contans one downlnk and one uplnk subframe n a TDD frame separated by the threshold, whch s dvded nto an nteger number of physcal slots (PSs). The downlnk subframe comes frst because t contans the bandwdth requests and transmsson nformaton drectly sent from SSs to the BS, whch forms a map for schedulng the uplnk resources among all the SSs. Fg. 2.4 The TDD frame structure n IEEE 802.6. The structure of the downlnk subframe usng TDD mode s llustrated n Fg. 2.5 [22], [24]. The downlnk subframe begns wth a frame start preamble used by the
PHY layer for synchronzaton and equalzaton. The preamble s followed by the frame control secton, contanng DL-MAP and UL-MAP statng the resource allocaton of the downlnk and uplnk. The DL-MAP specfes when the PHY layer transton occurs wthn the downlnk subframe. The followng porton carres the data, whch are transmtted to each SS usng a negotated burst profle. Due to the dynamcs of the bandwdth demand for the varetes of servces, the burst profles vary dynamcally from frame to frame. Thus the system can support dfferent levels of the data transmsson. In the case of TDD, a tme gap separates the downlnk subframe and the uplnk subframe. Preamble DIUC 0 DIUC DIUC 2 DIUC N Preamble ' DL-MAP UL-MAP Fg. 2.5 The TDD downlnk subframe structure n IEEE 802.6. 2.2.2 Schedulng Servce Schedulng servces have represented the data handlng mechansms supported by the MAC scheduler for the data transport on a connecton. To provde the servce parameters respectvely, the traffc management s necessary. The IEEE 802.6 standard dvdes all servces n four dfferent classes. Each group corresponds to a sngle servce class, whch s assocated wth a set of QoS parameters for quantfyng the aspects of ts behavor. Frstly, we outlne these four servce flows [22]: ) Unsolcted grant servce (UGS): It supports the constant bt rate (CBR) or fxed throughput connectons at perodc ntervals, such as T/E and voce over IP 2
(VoIP) whch needs to grant the constant bandwdth wthout any request. Ths servce can guarantee the data throughput and the latency. 2) Real-tme pollng servce (rtps): It s a real tme data stream comprsng varable bt-rate (VBR) data packets whch are ssued at perodc ntervals, such as the movng pctures experts group (MPEG) vdeo. Ths applcaton guarantees the mnmum reserved rate and the latency, whch are same as those of UGS. But the rtps has to request transmsson resources by pollng, whch means that the contenton and the pggyback are not allowed. 3) Non-real-tme pollng servce (nrtps): Ths servce s a delay-tolerant data stream consstng of varable-szed data packets, such as the fle transfer protocol (FTP). The mnmum data rate s requred and the bandwdth request by pollng s needed. 4) Best effort (BE): It does not provde any QoS guarantee, lke the emal or the short length FTP. There s no mnmum resources allocaton granted, where the occurrence of dedcated opportuntes s subject to the network load. The channel access mechansm of ths servce s based on the contenton. 2.2.3 Multuser Dversty Multuser dversty was ntroduced to deal wth tme-varyng fadng channels n multuser systems [9], [25]. In WMax, because of the multple users wth dfferent and tmevaryng channel condtons, there s a hgh probablty that one or several users have very good lnks wth the BS whle the others have bad ones. By adaptng to the channel condtons, the system throughput can be ncreased through the opportunstc schedulng. Ths s also referred to as the Adaptve Modulaton and Codng (AMC), whch can serve a large number of randomly dstrbuted users wth channel fadng ndependently. In ths case, the long term system throughput can be maxmzed by servng the users wth the best channel when transmttng. The scheduler structure of downlnk s depcted n Fg. 2.6. Although the proposed algorthm s presented n ths paper s for downlnk schedulng, ths dea can be extended to uplnk data transmssons. 3
Packet Classfer SS SS 2 SS N UGS rtps nrtps BE UGS rtps nrtps BE UGS rtps nrtps BE Downlnk Scheduler SS SS 2 SS 3 SS 4 SS N Fg. 2.6 Packet Scheduler structure 4
Chapter 3 An Overvew of Schedulng Algorthms Although there are a number of packet schedulng algorthms have been proposed for WMAX network [4-6], the desgn of those algorthms are challenged by supportng dfferent levels of servces, farness, and mplementaton complexty and so on. In ths chapter, we nvestgate three wdely adopted schedulng technques, Round Robn (RR), Proportonal Farness (PF), and Integrated Cross-layer schedulng (ICL) n wreless networks. 3. Round Robn (RR) RR s one of the smplest schedulng algorthms desgned especally for a tme sharng system, where the scheduler assgns tme slots to each queue n equal portons wthout prorty. Once a queue s served, t s not vsted agan untl all the other SSs n the system have been served. RR can provde a far resource access to each SS, and every queue s allocated wth the same porton of system resources regardless of the channel condton. However, the RR scheduler has the same bandwdth effcency as a random scheduler. Also, t cannot guarantee dfferent QoS requrements for each queue. 5
3.2 Proportonal Farness (PF) PF was proposed by Qualcomm Company, whch was realzed n the IS-856 standard for the downlnk traffc schedulng (also known as Hgh Data Rate (HDR)) [7-8]. The essental goals of ths packet schedulng scheme are to enhance the system throughput as well as provde farness among the queues under consderaton. Proportonal Farness schedulng s based on one prorty functon: µ r ( t ) ( t ) = R ( t ), where r ( t ) s the current data rate, R (t) denotes an exponentally smoothng average of the servce rate receved by SS up to slot t. Then the queue wth the hghest µ ( t) s served at tme slot t, where the average throughput of the queue s updated by R ( t + ) = ( ) R ( t ) ( ) r ( t ) T + c T, where T c s the tme constant for the c movng average. The average throughput of the queues that are not served at tme slot t s updated by the relatonshp: R ( t + ) = ( ) R ( t). In general, T c s assumed to be T c 000 slots (.66 seconds) n the CDMA-HDR system [7]. Takng a larger value of T c makes the perceved throughput less senstve to the short-term starvaton on the queue, where the scheduler may wat for a longer perod of tme for a user turnng back from a bad channel condton to a good one. When a huge number of users coexst n the system, we can obtan addtonal throughput gan by schedulng them to utlze the characterstcs of fast fadng channels, called mult-user dversty gan. Ths smple scheduler desgn enhances the overall throughput. In ths way, the preference metrc can be mplemented for achevng PF among the SSs [9-0]. Although PF s smple and effcent, t cannot guarantee any QoS requrement such as delay and delay jtter due to ts orgnal desgn for saturated queues wth non real-tme data servce. 3.3 Integrated Cross-layer Schedulng Both RR and PF cannot manage the resource allocaton and grants an approprate QoS per connecton. To resolve t, n the prevous work, [-3] proposed ntegrated QoS 6
control archtecture for IEEE 802.6 wreless systems. These schedulng schemes rely on dfferent algorthms to handle dfferent classes of servces for matchng ther QoS requrements. To have a comprehensve ntroducton, a representatve cross-layer schedulng algorthm wth QoS support by [] s brefed as follows. The scheduler bases on a prorty functon for each queue, where the prorty metrc of each queue s updated accordng to ts servce status and channel condton n the PHY layer. Thus, the scheduler can provde the prescrbed dverse QoS guarantees. Frst of all, the algorthm allocates a fxed number of tme slots for the UGS queues, whch s the requrement stpulated n the standard. The queues for real-tme Pollng Servce (rtps) are managed wth an Earlest Deadlne Frst (EDF) algorthm [4], whch s senstve to delay latency and relable for real-tme servces. An opportunstc scheme whch s smlar to the PF algorthm s deployed for the queues supportng non-real-tme Pollng Servce (nrtps), whle the queues for Best Effort (BE) traffc are managed based on a Best-Rate dscplne. In order to dfferentate the prorty of the four types of servces such that rtps > nrtps > BE, the class coeffcents are assgned to the queues of each servce type. The algorthm s mplemented accordng to the followng formulas. The total tme slots are allocated to Unsolcted Grant Servce (UGS) data streams as Nugs per frame. The resdual tme slots assgned to the rest QoS classes Nr = Nd Nugs, where N d s total tme slots n one frame. The prorty functon for connecton at tme slot t s defned as β φ ( t ) = β class class R ( t ) R F ( t ) N F ( t ) f F ( t ) < Where β class [0,] s the coeffcent accordng to the servce class, respectvely. Based on the prorty for dfferent QoS classes, rtps>nrtps>be, the coeffcent can be set under the constrant β > rt β nrt > β BE. R (t) s the number of bts can be carred by one symbol at frame t va AMC, whch s determned by the channel condton. R N denotes the maxmum number of bts n one symbol can be reached. f 7
For each rtps connecton, F(t) s an ndcator of the delay satsfacton. F (t)= T b - W (t), wth R N denotng the delay bound and W ( t) [0, T ] denotng the longest packet watng tme. For smplfyng the formula, we do not consder the guard tme here. Ths prorty functon normalzes φ ( t) [0, β ]. When F (t) <, whch means the packets n the queue rt should be sent mmedately to avod the packet drop, the hghest values β has been set. rt For each non real tme connecton, F (t) s the raton of the average transmsson rate to the mnmum reserved rate. ηˆ ( t) F ( t) =, where ηˆ ( t) s estmated by η ˆ η ( t + ) = ( ) ˆ η ( t ) + ( ) r ( t ), r ( t ) s the transmsson rate at tme t. At ths tme, F (t) T c T c s an ndcator of the data rate satsfacton. So f F (t) <, the packets of the th stream should be sent to meet the rate requrement. The upper bound of prorty value for nrtps s β. nrt Because there s no QoS requrement for BE connectons, the prorty functon for a BE R ( t ) connecton s φ ( t) = β. BE φ ( t) only depends on the normalzed channel qualty R regardless of the delay or rate performance. n Ths scheme provdes a dverse QoS support for multple connectons. However, the author cannot offer how to set the upper bound of β, rt β, nrt β, tme slots number reserved for UGS Nr BE and delay bound T to get the optmal performance of the system. The equvalence of dfferent prorty functons for four types of servces has not been proved. The scheduler s hard to be practcally deployed due to ts hgh mplementaton complexty. 8
Chapter 4 Scheduler Desgn 4. System Model We propose a novel scheduler n ths chapter, whch s desgned for a fxed PMP WIMAX system. Only one BS n the network serves all the SSs, therefore the nter-bs nterference can be neglected. The downlnk channel s shared by all SSs n a tme dvson multplexng manner, where a downlnk scheduler s deployed at the BS to schedule the transmssons correspondng to the queues. The transmsson occurs wthn a fxed-szed tme frame. Only the selected queues can be severed wthn the frame. The ar nterface specfcaton s OFDM whch employs a fast Fourer transform (FFT) of sze 256. All carrers are assgned to one queue for the data transmsson n a tme slot Over a wreless fadng channel, AMC descrbed n 2..2 s employed at the PHY layer. The BS exactly knows the channel state nformaton of all the SSs at each tme frame. As specfed n the standard, ndvdual SS measures the SNR and feedbacks the nformaton to the BS scheduler n each tme frame. The BS receves the feedback sgnals from all the SSs to collect the current channel status. Wth the perfect channel state nformaton, the BS scheduler makes the resource allocaton decson for queues and selects the sutable adaptve codng and modulaton on each traffc channel. 9
4.2 Channel Fadng Model An OFDM system transfers a broadband sgnal nto parallel narrowband sub-channels, thus the frequency selectve fadng can be overcame. Therefore, we adopt the general Raylegh channel model that s sutable for flat-fadng channels as well as frequencyselectve fadng channels encountered wth OFDM [26-27]. The average sensed SNR of SS k can be expressed as SNR P k k =, where n Pn P s the background nose varance [28]. The recevng power of SS k s gven by P = h P, where P t s the total transmtter power of the BS and h k s the channel gan, whch reflects the effects of several physcal phenomena ncludng scatterng, obstacles, and multpath propagaton. In further detal, α the channel gan from the BS can be wrtten as h = cdk Sk mk, where c s a constant ncorporatng the transmsson and recevng antenna gans, d k s the dstance from the BS to user k, α s the path loss exponent, S k s a random varable for the shadow fadng effect, whch s known to follow the log-normal dstrbuton wth zero-mean and varance 2 σ s (db) n the log-scale. The multpath fadng effect m k s modeled as an exponental random varable wth a mean.0, whch represents the Raylegh fadng channel. We also defned the medan SNR at the cell edge, ρ, to represent the nose level of the wreless envronment consdered, whch has been appled n prevous research [29] ρ = cd P P, where D s the radus of the SS. α t / n 2 α 2 k k n t k n ρ k SNR = P / P = P / h P = D h / c SNR = D d S m = D d S α 2 α k ρ( / k ) k k ρ( / k ) k For a Raylegh fadng channel, the receved SNR s modeled by an exponental random k 2 k varant [30]. The probablty densty functon s gven by x g( x) = exp( ) λ x λ 0, where s the average receved SNR. λ k t 20
4.3 Proposed Algorthm Schedulng 4.3. Prorty Functon For the best effort traffc, the well known PF s an attractve dscplne. In ths secton, the proposed schedulng algorthm, called Adaptve Proportonal Farness (APF) schedulng, s ntroduced, whch ams to extend the PF schedulng to the real tme servce and provdes dverse QoS requrements, The schedulng scheme s based on the Grant Per Type-of Servce (GPTS) prncple, whch ams to dfferentate the delay performance of each queue. A novel prorty functon s devsed for all the QoS guarantee queues, ncludng UGS, rtps, and nrtps, for allocatng tme slots on the queues wth the hghest prorty value. At the tme nterval t, the prorty functon for queue s defned as: C W ( t ) ( t ) = (4.) R ( t ) / ( K ( t ) M ) where M s the mnmum rate requrement, K ( t ) s the number of connectons of the th queue. W (t) s the transmsson capacty at tme t, whch s determned by the channel qualty. As n Table 2, W (t) s smplfed to the number of bts that can be carred by one symbol va AMC. Each queue corresponds to one QoS requrement class, respectvely. The estmated average throughput of queue (denoted as R ( t ) ) s updated by the smple exponental smoothng model as follows: R ( t ) = ( ) R ( t ) ( ) r ( t ) T + T (4.2) where r ( t ) denotes the current throughput and T s a predefned system parameter assocated wth each class. Rather than makng T constant n the PF scheme, we assgn dfferent values of T to each type to perform the dfferent applcatons. In ths thess, we emphasze on explanng the proposed preference metrc avalable to support dverse servces wth dfferent QoS requrements. We assume the varaton of channel condtons can allow us to select the same T for the same servce type. The SSs, whch are close to the BS, wll not sacrfce the transmsson of the others. 2
The key parameter of ths algorthm s T n Eq. (4.2). The role of T s to dstngush the prorty for dfferent types of servces. The avalable analyss of Eq. (4.2) has been shown n [3]. Wth a smallt, the preference metrc of queue fluctuates sgnfcantly, makng queue beng vsted frequently. Ths feature s crtcal to dstnct the queue wth the delay constrant. For example, n the case T =2, accordng to Eq. (4.2), f the queue has not been served by the scheduler at prevous n tme frames, R ( t) = R ( t n) n. R (t) 2 decreases dramatcally, whch lead queue to wn the transmsson opportunty at the current tme frame more lkely. In contrary, wth a large T, (for nstance, when T =00) R (t) decreases slghtly although the queue has not been vsted for a few tme frames. Thus, the channel qualty W (t) s the man determnng factor for the prorty of each queue. In ths case, the system throughput s enhanced and effcent bandwdth utlzaton s acheved. However, the scheme s not susceptble to the latency of the connectons. Wth such a desgn, the task s fndng a proper T to acheve the balance of delay and throughput for each class accordng to QoS requrements. Quantty R (t)/ K (t) specfed n Eq. (4.) normalzes the throughput of each connecton n the th queue. Whle R (t)/(k (t)m ) s an ndcator of data rate satsfacton, whch also reflects the delay satsfacton for real tme connectons. For a real-tme servce, R (t) evaluated wth a small T s senstve to the watng tme of the data n queue. Large value of R (t) ndcates hgh degree of delay satsfacton, whch leads to low prorty. Therefore, a varety of QoS demands for real-tme and non-real-tme applcatons are unfed to R (t)/(k (t)m ), whch plays an essental role n reflectng the nstantaneous bandwdth requrements of queue. Jontly consderng the effect of the current channel condtons and the transmsson satsfacton of the prevous tme frames, the proposed preference metrc not only keeps 22
the good advantages of PF scheme, but also dfferentates the servces wth dverse QoS requrements by selectng an approprate set of T. 4.3.2 Tme Slot Allocaton A cyclc MAC scheduler operates on one frame bass. There are n PSs n a gven frame, where n s determned by the system parameter settng. The average throughput of each queue s tracked by ts exponental movng average. At the begnnng of each tme frame, the BS calculates the preference metrc, whch s defned as the Eq. (4.). The queue wth the maxmum preference metrc wll be selected for transmsson at the next comng PS. Upon each vst, all the data sub-carrers are assgned to the correspondng queue and all data n the served queues are transmtted except that the remanng capacty of the current frame s not large enough to fully accommodate all data. In ths way, the throughput of all connectons s guaranteed. When the capacty of the current slot s larger than the data n the served queue, the data n the BE queue whch belongs to the same SS can be transmtted usng the remanng capacty. Once all the PSs n the current frame are exhausted, the rest queues that are not yet processed must be served later. Then, the schedulng process wll repeat n the next frame. 23
Chapter 5 Analyss and Smulaton Results In Chapter 4, we have proposed the new schedulng scheme APF. The performance of the APF scheme s further nvestgated analytcally and va computer smulatons wth MATLAB n ths chapter. A WMAX network model wth a sngle BS s developed as the system model for the analyss and smulatons. Frst, we examne the capablty of the APF scheme to dfferentate the QoS of dfferent servce classes such as the average delay. Second, the achevable system throughput of the APF scheme s evaluated. After that, the system performance n the presence of traffc load varaton s consdered. Also, we examne the scalablty of the APF scheme to the system sze n terms of the number of accommodated SSs. Last, the performance of the APF scheme s compared wth three other abovementoned schemes. 5. Performance Analyss We do not consder the channel condtons to smplfy the analyss of the effect of T. Ths assumpton s reasonable when T s rather small. Thus, the preference metrc can be C (t) rewrtten as Eq. (5.). C ( t ) = K M R ( t ) (5.) where R (t) denotes the average throughput at tme frame t. When queue has not been served n the contnuous N tme frames, R (t) s updated by Eq. (5.2). N R ( ) ( ) t = R ( t N ) (5.2) T 24
Then, C K M K M C ( t N ) ( t ) = = = R ( ) N t N ( ) R ( t N ) ( ) T T We assume that the values of the preference metrc of the wnnng queue converge to a constant when the system s stable, proved by [33]. An approxmate relatonshp between the nter-servce tme (denoted as N ) and T s obtaned as: C ( t N ) C ( t N ) N N ( ) ( ) T T, The approxmate relatonshp between N and N or D and D (D denotes the average delay of queue ) s gven by Eq. (5.3). N N D = = D lo g ( ) + δˆ, ( t ) T lo g ( ) T, (5.3) where ˆ lo g C ( t N ) lo g C ( t N ) δ, ( t ) =. N The accurate computaton for C (t) s complcated due to the mpact of many factors, such as the traffc load and arrval rate of connectons. It must be noted that the value of ( / T ) s so small that δ ˆ ( ), t n the Eq. (5.3) cannot be gnored. Thus, t s qute dffcult to establsh a mappng between T and D. Here, a smplfed equaton reformulated as Eq. (5.4) provdes an approxmate mappng relatonshp. N N D = = D lo g ( T ) lo g ( T ), (5.4) 25
5.2 Smulaton Results 5.2. System Parameter Settng and Assumpton 5.2.. Channel Model As mentoned n secton 4.2, all the SSs are assumed to have ndependent Raylegh fadng channels. The average receved SNR of each SS s derved by α SNR = ρ( D / d ) S k k k Each SS radus D s set to (km). The path loss exponent α s 2, the standard devaton of large scale fadng σ s s set to 8 (db) and the medan SNR at the cell edge ρ s 0 (db), referred to [28]. In smulaton, 0 SSs are located. The dstances away from the BS are lsted n Table 3. Table 3 Dstances from SS to the BS Index of SSs 2 3 4 5 6 7 8 9 0 Dstance (km) 3.8 4.5 5 4.5 6 3.8 5 4 5 6 At each tme frame, the receved SNR s a random varable wth an exponental dstrbuton. The optmum modulaton and codng rate for the correspondng channels are chosen from Table 2. 5.2..2 System Model The whole bandwdth s assumed to be 20MHz and duraton of one tme frame s 2.5ms. In the OFDM 256 FFT mode, the duraton of each OFDM symbol s 2.6µs and one PS s composed of four OFDM symbols. Thus, a sngle 2.5ms tme frame conssts of around 50 OFDM PSs, 20 PSs of them algned to downlnk data transmsson. The duraton of each frame s so short that t s reasonable to omt the channel fluctuaton durng one frame for a fxed wreless system. In each SS, there are four queues correspondng to four traffc classes. Each queue has an nfnte backlog of data. In the smulaton, all traffcs arrve at the begnnng of each frame. The generator sources K of queue s constant. For each 26
UGS, rtps and nrtps connecton, we assume that the arrval process to the queue follows a Posson dstrbuton wth a gven arrval rate, whch s gven n Table 4. For BE servce, we assume the queues are always saturated. Table 4 Input servce flow of each SS Servce Type Average arrve rate Mn.reserved rate of each connecton (kbps) (kbps) UGS 9.6 8 rtps 80 64 nrtps 5 4 5.2.2 Smulaton Results 5.2.2. The Performance of APF A. Capablty of Dfferentaton and the delay performance of APF We frst nvestgate the effect of dfferent T on the servce prorty to verfy the capablty of APF to dfferentate the servce classes. A large value of T makes APF scheme roughly behave as PF scheme. The system throughput s hgh, but the data n the queue may experence large delay. For a real-tme applcaton, conformng delay constrant s more crtcal than mprovng the long-term throughput. However, for a non-real-tme applcaton, the major concern s the long-term throughput. Accordng to the delay constrant of each traffc type, the values of T should be set as T UGS < T rtps < T nrtps. Two examples are llustrated n Fg. 5. to analyze the mpact of T, whch present the average prorty value of each class n the full load system. 27
4 x 0-3 x 0-4 3.5 3 2.5 UGS rtps nrtps 3 Prorty value 2 Prorty value 2.5 0.5 UGS rtps nrtps 0 0 0 20 30 40 50 60 70 80 90 00 Frame 0 0 0 20 30 40 50 60 70 80 90 00 Frame (a) (b) Fg. 5. Comparson for Prorty value under dfferent T sets (a) T of UGS, rtps, nrtps are [2, 4, 8] (b) T of UGS, rtps, nrtps are [200, 400, 800] Accordng to Eq. (4.), the peak ponts n the above fgures mean queues are served. As shown n Fg. 5., wth T ncreasng, the prorty values are not senstve to the parameter T. Although the rule T UGS < T rtps < T nrtps s enforced and the ratos TUGS : TrtPS : T nrtps are equvalent n two experments, t s observed that, n Fg. 5. (b), the fluctuaton characterstcs of the prorty values for all types of servces are consstent, followng the channel condton varaton. In ths case, the prorty of the connecton manly depends on ts channel qualty. Correspondng to Fg. 5. (b), the fgures of the related average delay are presented n Fg. 5.2 (b). Obvously, three types of queues cannot be dfferentated n terms of the average delay when each T s large. In contrast, Fg. 5. (a) shows the UGS queues have the smallest average delay whle the nrtps queues have the largest average delay. Therefore, the selected T for UGS and rtps connectons should be so small that the prorty values are susceptble to the watng tme of data. Whle the selected T for nrtps should be large enough that the rankng s manly determned by the channel qualty. It s shown that APF can provde the servce dfferentaton among multple servce classes by selectng a proper set of T. 28
To further verfy the effects of T, more smulatons are conducted, shown n Fg. 5.2 (c)- (f). We change one of the T n a set and all other parameters are kept same. By the comparson of the three pars, (a) and (c), (d) and (e), and (d) and (f), t s observed that the average day ncreases wth the related T ncreasng and the average delay of the rest two servce types decreases. It s clear that the average delay s coupled wth T. Ths result makes T can be manpulated to provde servce dfferentaton n terms of the average delay for each class. 2.4 Frame 5 Frame 2.2 4.5 2.8 UGS rtps nrtps 4 3.5 Average delay.6.4.2 Average delay 3 2.5 2 UGS rtps nrtps.5 0.8 0.6 0.5 0.4 2 3 4 5 6 7 8 9 0 Traffc loadng 0 2 3 4 5 6 7 8 9 0 Traffc loadng (a) (b) 29
4 Frame 2.5 Frame 3.5 3 UGS rtps 2 UGS rtps nrtps 2.5 nrtps Average delay 2.5 Average delay.5 0.5 0.5 0 2 3 4 5 6 7 8 9 0 Traffc loadng 0 2 3 4 5 6 7 8 9 0 Traffc loadng (c) (d) 3 Frame 2.4 Frame 2.2 2.5 2 2 UGS rtps nrtps.8 UGS rtps nrtps.6 Average delay.5 Average delay.4.2 0.8 0.5 0.6 0 2 3 4 5 6 7 8 9 0 Traffc loadng 0.4 2 3 4 5 6 7 8 9 0 Traffc loadng (e) (f) Fg. 5.2 Comparson for Average delay under dfferent T sets (a) T of UGS, rtps, nrtps are [2, 4, 8], (b) T of UGS, rtps, nrtps are [200, 400, 800] (c) T of UGS, rtps, nrtps are [2, 4, 00] (d) T of UGS, rtps, nrtps are [4, 0, 50] (e) T of UGS, rtps, nrtps are [2, 0, 50] (f) T of UGS, rtps, nrtps are [4, 20, 50] 30
The followng s a case study on examnng Eq. (5.4). As seen from Fg. 5.2, Fg. (a), T = [2, 4, 8] and Fg. (d), T = [4, 0, 50], are very smlar accordng to both the upward trend and the value of the average delay. Among Fg. 5.2 (a)-(f), the values of ( ) log / T : log( / T ) : log( / T ) of two sets are closest. To further verfy the ugs rtps nrtps Eq. (5.4), another two sets are selected, whch are lsted n Table 5 wth the correspondng calculated values. Wth the qute equal rato of ( ) log / T :log( / T ):log( / T ), Fg. 5.2 (a), (g) and (h) have shown three almost ugs rtps nrtps dentcal fgures, whch confrm us to approxmate estmate T wth the gven delay constrant by Eq. (5.4). Table 5 T and the Correspondng Analyss Values T T2 T3 log(-/t):log(-/t2) log(-/t):log(-/t3) Fg 2 4 8 2.4 5.9 Fg. 5.2(a) 4 9 9 2.44 5.32 Fg. 5.2(g) 8 20 40 2.6 5.27 Fg. 5.2(h) 2.4 Frame 2.4 Frame 2.2 2.2 UGS 2 rtps nrtps 2 UGS rtps.8.8 nrtps Average delay.6.4.2 e lay d ge A vera.6.4.2 0.8 0.8 0.6 0.6 0.4 2 3 4 5 6 7 8 9 0 Traffc loadng 0.4 2 3 4 5 6 7 8 9 0 Traffc loadng (g) (h) Fg. 5.2 Comparson for Average delay under dfferent T set (g) T of UGS, rtps, nrtps are [4, 9, 9] 3
(h) T of UGS, rtps, nrtps are [8, 20, 40] 70 Mbps 70 Mbps 60 60 50 50 Throughput 40 30 Throughput 40 30 20 20 Capacty 0 Total Throughput UGS&rtPS&nrtPS Throughput 0 Capacty Total Throughput UGS&rtPS&nrtPS Throughput 0 2 3 4 5 6 7 8 9 0 Traffc loadng 0 2 3 4 5 6 7 8 9 0 Traffc loadng (a) (b) 70 Mbps 70 Mbps 60 60 50 50 Throughput 40 30 Throughput 40 30 20 20 0 Capacty Total Throughput UGS&rtPS&nrtPS Throughput 0 Capacty Total Throughput UGS&rtPS&nrtPS Throughput 0 2 3 4 5 6 7 8 9 0 Traffc loadng 0 2 3 4 5 6 7 8 9 0 Traffc loadng (c) (d) Fg. 5.3 Comparson for Throughput under dfferent T sets (a) T of UGS, rtps, nrtps are [2, 4, 8], (b) T of UGS, rtps, nrtps are [200, 400, 800] (c) T of UGS, rtps, nrtps are [2, 4, 00] 32
(d) T of UGS, rtps, nrtps are [4, 0, 50] B. The system throughput of APF In ths secton, extensve smulaton s conducted to nvestgate the relatonshp between T and the system throughput. Based on dfferent T sets as shown n Fg 5.2, the correspondng capacty, the total throughput, and the cumulatve throughput of UGS, rtps and nrtps are shown n Fg. 5.3. The system capactes are dentcal n all fgures n Fg. 5.3 due to an deal status where the selected queues are all n the best channel condton. The total throughput s defned as an effectve data rata transmtted from the BS to all the SSs, whch s depcted by the blue lne. The data streams nclude UGS, rtps, nrtps and BE. Snce the selected queue s served exhaustvely, the cumulatve throughput of UGS, rtps and nrtps shown by the green lne should be equal to the sum of ther arrval data rate. It s observed that the mnmum throughput constrant of each connecton can be exactly met, whle the bandwdth effcency s much mproved compared wth that usng RR and ICL. However, the choce of T does not make much dfference n the total throughput. Ths characterstc makes the manpulaton of T only base on the delay constrant of each servce type for QoS provsonng. Meanwhle, t shows that APF takes the advantage of opportunstc schedulng schemes, enhancng the system throughput regardless of the value of T. C. The mpact of the traffc load to APF Next, we wll nvestgate the mpact due to the varaton of the traffc load. Fg. 5.2 and Fg. 5.3 show the relaton between the average delay and the throughput wth the dfferent traffc load. The Fg. 5.2 ndcates that APF does not assgn much more prvleges to UGS and rtps when the system s n the lght traffc. The dfferences of the delay performance for the varous classes are not apparent. Because when the capacty resource s bg enough, all connectons have the chance to be served to satsfy the transmsson demands. Consequently, all queues acheve the small average delay. In Fg. 5.3, when the system s n the lght traffc, the throughput of large T s slghtly hgher 33
than that of the small T. It s clear that the performance dfference among the three servce types s not sgnfcant n a lghtly loaded system. In ths case, the advantages of APF are not promnent. In addton, wth enough resources, most conventonal schedulng schemes can work well, such as RR, PF and etc. Whle n a heavy traffc system, the lmted resource has to be carefully assgned to each connecton. The schedulng scheme s especally crtcal to acheve the satsfed performance for each servce. Under such crcumstance, APF exhbts ts advantage obvously that the scheduler can gve the preference to the servces n terms of ther delay constrants. D. The mpact of the number of SSs to APF Both pctures n Fg. 5.2 and Fg. 5.3 present the algorthm performance when addng new connectons to decrease the avalable capacty for each flow. The performance of the servce classes wth the low prorty degrades pror to that of the servce classes wth the hgh prorty. Scalablty s acheved. We keep trackng the scalablty of the proposed scheme by addng several new SSs, but decrease the number of connectons to balance the system n a full traffc load. In ths set of experments, the numbers of SSs are vared from 6 to 40, but the number of PSs n one tme frame keeps unchanged. Intutvely, the vsted tmes for each queue decrease n the fxed tme duraton wth the number of SSs ncreasng and each queue has to wat longer for beng vsted. Thus, the average delay for each queue ncreases. 34