The Pennsylvania State University. The Graduate School. College of Engineering

Transcription

1 The Pennsylvana State Unversty The Graduate School College of Engneerng ADAPTIVE RESOURCE ALLOCATION FOR D-TDD SYSTEMS IN WIRELESS MULTIMEDIA NETWORKS A Thess n Electrcal Engneerng y Jungnam Yun 2004 Jungnam Yun Sumtted n Partal Fulfllment of the Requrements for the Degree of Doctor of Phlosophy May 2004

2 The thess of Jungnam Yun was revewed and approved* y the followng: Mohsen Kavehrad Wess Char Professor of Electrcal Engneerng Thess Advsor Char of Commttee John J. Metzner Professor of Electrcal Engneerng John F. Doherty Assocate Professor of Electrcal Engneerng Davd J. Mller Assocate Professor of Electrcal Engneerng Natarajan Gautam Assocate Professor of Industral Engneerng W. Kenneth Jenkns Professor of Electrcal Engneerng Head of the Department of Electrcal Engneerng *Sgnatures are on fle n the Graduate School

3 ABSTRACT In accordance wth the demand for roust and spectrally effcent communcatons systems desgn for multmeda communcaton wth heterogeneous traffc n fadng channels, varous adaptve approaches have een developed, recently. Ths dssertaton examnes how adaptve technologes can e used n real systems n order to rng relale and spectrally effcent connectons etween wreless communcaton lnks. It addresses two dfferent types of adaptaton: (1) channel adaptaton wth adaptve modulaton /codng, adaptve antennas, and automatc repeat request (ARQ) schemes; (2) andwdth adaptaton to the dynamc unalanced network traffc wth dynamc tme dvson duplex (D-TDD) scheme. In order to make the vulnerale wreless channel relale and spectrally effcent, we adopt adaptve modulaton/codng and ARQ schemes. We ntroduce an analytcal method, whch uses a fnte-state Markov chan (FSMC) as an error model, for estmatng the performance of adaptve modulaton systems (AMS) comned wth ARQ schemes n correlated slow fadng channels. For the purpose of the throughput performance evaluaton of wreless packet networks, fadng channels have een assumed to have ndependent and dentcally dstruted (..d) statstcs. Ths assumpton of channel model may e sutale to represent fast fadng channels. However, n slow fadng channels, error rates of consecutve packets are hghly correlated, and we cannot smply assume an ndependent error process for the purpose of performance evaluaton. We propose a mult-state Markov error model for AMS n correlated fadng channels, whch s also descred y a Fnte-State Markov Chan (FSMC), and we present throughput estmaton methods for AMS comned wth ARQ, usng the proposed Markov error model. For multmeda servces, network traffc ecomes heterogeneous; the network traffc s homogeneous for voce traffc only. It has een known that dynamc-tdd (D- TDD) scheme utlzes the spectrum more effcently, adjustng the rato of uplnk and

4 v downlnk avalale tmeslots adaptvely accordng to the traffc pattern. However, havng dfferent swtchng ponts n each cell causes severe co-channel nterference etween ase statons. We present a statstcal assessment of co-channel nterference of D-TDD scheme when swtched eam sector antennas are deployed at ase statons, and we also present a cooperatve co-channel nterference avodance algorthm n TDMA/D-TDD wreless cellular systems. Fnally, wth these channel and andwdth adaptatons, we propose adaptve resource allocaton schemes consderng jont parameters n the physcal (PHY) layer and the medum access control (MAC) layer. We consder cases wthout co-channel nterference and wth co-channel nterference, denotng as sngle and mult- cluster cases. For the sngle cluster case, we frst allocate andwdths for uplnk and downlnk ased on the requested andwdths for oth lnks. Secondly, we allocate modulaton format to each call ased on the estmated channel state; ths procedure makes some more channels avalale n the frame for the calls watng n the queue. For the mult-cluster case, n addton to resource allocaton procedures for the sngle cluster case, we have to allocate tme slots to each call approprately n a way of avodng severe co-channel nterference. By smulaton results, we show that more calls (or users) can e servced n D-TDD systems than fxed-tdd (FTDD) systems wth (or wthout) adaptve modulaton.

5 TABLE OF CONTENTS v LIST OF FIGURES...v LIST OF TABLES...x LIST OF ACRONYMS...x ACKNOWLEDGEMENTS...x Chapter 1 INTRODUCTION Motvaton Overvew of Adaptve Technques Adaptve Modulaton and Codng Adaptve ARQ Schemes Dynamc TDD Systems Organzaton of the Dssertaton...10 Chapter 2 FINITE STATE MARKOV CHANNEL MODEL Introducton Fnte-State Markov Channel (FSMC) Models Smulaton of Fadng Channel Models Clarke s Model Jakes Model Numercal Results Conclusons...25 Chapter 3 MARKOV ERROR MODEL FOR BLOCK TRANSMISSIONS IN ADAPTIVE MODULATION SYSTEMS Introducton Error Models for Fadng Channels Two-state Error Model Mult-State Error Model Numercal Results Conclusons...37 Chapter 4 ADAPTIVE RATE TRANSMISSIONS COMBINED WITH ARQ Introducton System Confguratons Throughput Performance of AMS comned wth ARQ Adaptve Modulaton comned wth Go-ack-N ARQ Adaptve Modulaton comned wth Selectve-repeat ARQ...50

6 4.4 Cross-Layer Optmzaton Numercal Results Conclusons...56 Chapter 5 DYNAMIC TIME DIVISION DUPLEX SCHEME FOR HETEROGENEOUS TRAFFIC Introducton System Confguratons Frame Format Traffc Model Co-channel Interference Co-operatve TSA Algorthm for D-TDD Systems Numercal results Conclusons...73 Chapter 6 ADAPTIVE RESOURCE ALLOCATION FOR TDD/TDMA SYSTEMS WITH HETEREGENEOUS TRAFFIC Introducton System Confguratons Frame Format Channel Model Adaptve Modulaton Call (Connecton) Model Resource Allocaton for TDD/TDMA Systems Sngle-cluster Case wth Fxed Rate Transmssons Sngle-cluster Case wth Adaptve Rate Transmssons Mult-Cluster Case wth Fxed Rate Transmssons Mult-Cluster Case wth Adaptve Rate Transmssons Performance Evaluaton Traffc Model Comparson wth Fxed TDD (F-TDD) Outage Proalty Numercal Results Conclusons Chapter 7 CONCLUSIONS AND FUTURE WORK Summary of Results Future Research Drectons Lst of Contrutons Blography v

7 LIST OF FIGURES v Fgure 1-1: An example of roadand wreless access (BWA)...2 Fgure 1-2: Communcaton degrees of freedom...3 Fgure 2-1: The fnte-state Markov channel model...15 Fgure 2-2: Clarke s fadng smulator usng quadrature ampltude modulaton wth aseand Doppler flter...19 Fgure 2-3: Fadng envelope levels generated y Clarke s model for dfferent vehcle speeds wth f c =1.9GHz: (a) v = 5km/h (f D = 8.796), () v = 20km/h (f D = )...20 Fgure 2-4: Jakes fadng smulator wth a sum of snusods...21 Fgure 2-5: Fadng envelope levels generated y Jakes model for dfferent vehcle speeds wth f c =1.9GHz: (a) v = 5km/h (f D = 8.796), () v = 20km/h (f D = )...22 Fgure 2-6: Autocorrelaton of the real and magnary components of the receved complex envelope for Jakes smulator and Clarke's smulator...23 Fgure 2-7: Markov channel model; v=5km/h, R=100k/s, N F =424 symols, τ =3T F, f m T F =0.0373, N=11; (o) smulaton, (x) analytcal model...24 Fgure 2-8: Markov channel model; v=5km/h, R=1M/s, N F =424 symols, τ =8T F, f m T F =0.0037, N=53; (o) smulaton, (x) analytcal model...25 Fgure 3-1: Glert-Ellot model: G = good state, B = ad state...28 Fgure 3-2: Extended Markov error model...30 Fgure 3-3: Markov error model; v=5km/h, R=100k/s, N F =424 symols, τ =3T F, f m T F =0.0373; (o) smulaton, (x) analytcal model...35 Fgure 3-4: Markov error model; v=5km/h, R=1M/s, N F =424 symols, τ =3T F, f m T F =0.0037; (o) smulaton, (x) analytcal model...36 Fgure 4-1: Confguraton of adaptve modulaton systems...40 Fgure 4-2: Exemplary frame structure for adaptve modulaton systems...41 Fgure 4-3: Curve fttng test to fnd parameters: a, g, γ p as shown n Tale 4-1. (- ) exact FER curves, (*) approxmated FER curves, and (--) exact BER curves

8 for {BPSK, QPSK, 8-QAM, 16-QAM, 32-QAM, 64-QAM, 128-QAM, 256- QAM}...43 Fgure 4-4: Go-ack-N ARQ example...47 Fgure 4-5: A typcal sequence example for Go-ack-N ARQ...48 Fgure 4-6: Selectve-repeat ARQ example...50 Fgure 4-7: Throughput performance for v = 5km/h, R=1M/s, N F =424 symols, f m T F = Fgure 4-8: Throughput performance for v = 20km/h, R=1M/s, N F =424 symols, f m T F = Fgure 4-9: Throughput performance for v = 10km/h, R=100k/s, N F =424 symols, f m T F = Fgure 4-10: Throughput performance for v = 20km/h, R=100k/s, N F =424 symols, f m T F = Fgure 4-11: Fndng an optmal target FER for maxmum throughput performance...56 Fgure 5-1: D-TDD/TDMA frame structure...61 Fgure 5-2: Example of co-channel nterferers...64 Fgure 5-3: Geometrcal model of co-channel nterference when swtched eam sector antennas are deployed at BS stes...65 Fgure 5-4: Co-operatve TSA algorthm - two cells example...68 Fgure 5-5: CDF of SIR when omn-drectonal antennas are deployed at BS...70 Fgure 5-6: CDF of SIR when sector antennas are deployed at BS (efore and after TSA)...71 Fgure 5-7: Aggregate SIR wth the 6 extra tme slots for 8 (18) sector antennas at BS...72 Fgure 5-8: Aggregate SIR wth the 12 extra tme slots for 8 (18) sector antennas at BS...73 Fgure 6-1: Frame format of D-TDD system wth two classes of servces...77 Fgure 6-2: An example of adaptve modulaton for BWA...79 v

9 Fgure 6-3: Numer of calls for class A and class B...84 x Fgure 6-4: Strong co-channel nterference from ase statons n co-channel cells...86 Fgure 6-5: TSA wth two cells...87 Fgure 6-6: Normalzed spectral effcency for sngle-cluster case wth fxed modulaton...92 Fgure 6-7: Numer of admtted calls for sngle-cluster case wth fxed modulaton...93 Fgure 6-8: Normalzed spectral effcency for sngle-cluster case wth adaptve modulaton...94 Fgure 6-9: Numer of admtted calls for sngle-cluster case wth adaptve modulaton...95 Fgure 6-10: Normalzed spectral effcency for mult-cluster case wth fxed modulaton...96 Fgure 6-11: Numer of admtted calls for mult-cluster case wth fxed modulaton...97 Fgure 6-12: Normalzed spectral effcency for mult-cluster case wth adaptve modulaton...98 Fgure 6-13: Numer of admtted calls for mult-cluster case wth adaptve modulaton...99

10 LIST OF TABLES x Tale 4-1: Parameters for approxmated FER equatons; N C = 40, N P = 384 symols...43 Tale 4-2: Example of threshold levels and average FER; average SNR, γ 0 =18dB target FER=0.1, and τ =3T F...44 Tale 6-1: Example of call model...81

11 LIST OF ACRONYMS x AFSD AMS ARQ AWGN BER BS BWA CCI CDF DSL D-TDD FDD FEC FER FSMC F-TDD GBN-ARQ HFC HPBW Average Fade State Duraton Adaptve Modulaton System Automatc Repeat request Addtve Whte Gaussan Nose Bt Error Rate Base Staton Broadand Wreless Access Co-channel Interference Cumulatve Dstruton Functon Dgtal Suscrer Loop Dynamc Tme Dvson Duplex Frequency Dvson Duplex Forward Error Correcton Frame Error Rate Fnte State Markov Chan Fxed Tme Dvson Duplex Go-Back-N ARQ Hyrd Fer Cale Half Power Beam Wdth

12 ..d. ITU LMDS LOS MAC NLOS PDA PER RCPC SIR SNR SR-ARQ TDD TDMA TSA Independent and Identcally Dstruted Internatonal Telecommuncatons Unon Local Multpont Dstruton Servce Lne of Sght Multple (Medum) Access Control Non Lne of Sght Personal Dgtal Assstant Packet Error Rate Rate Compatle Punctured Convolutonal Sgnal to Interference Rato Sgnal to Nose Rato Selectve Repeat ARQ Tme Dvson Duplex Tme Dvson Multple Access Tme Slot Allocaton x

13 ACKNOWLEDGEMENTS x Prase e to God, for He s good and Hs love endures forever. Wthout Hs Grace and Love, the accomplshment of ths thess would never have een possle. I would lke to acknowledge many ndvduals for ther support and care durng my doctoral studes. Frst, I would lke to thank my thess advsor, Professor Mohsen Kavehrad, for gvng me the opportunty to e a part of hs research group and for provdng me the rght alance of gudance and ndependence n my research. I am greatly ndeted to hm for hs advce oth n techncal and non-techncal matters, and for the fnancal support over the past four years. I am also grateful to Professor John J. Metzner, John F. Doherty, Davd J. Mller, and Natarajan Gautam for servng on my dssertaton commttee and for ther varous suggestons relatng to my research. I would lke to thank the former and present memers of the Center for Informaton and Communcatons Technology Research for ther frendshp and advce. My specal thanks go to my parents for ther love, trust, and prayers. I am also grateful to my parents-n-law for ther encouragement and prayers. My fnal acknowledgement goes to my wfe Jung for her dedcaton and sacrfce. Ths dssertaton would have never een completed wthout her presence esde me.

14 Chapter 1 INTRODUCTION 1.1 Motvaton Wth the remarkaly ncreasng demand for roadand wreless access such as hgh-speed Internet/we access, fast, deployale, and cost-effectve solutons have een desred. To fulfll ths request, dgtal suscrer loop (xdsl), cale modems over hyrd fer cale (HFC), and optcal fers have een ntroduced. However, due to the dstance restrcton of DSL and expensve cost to equp every house wth cales or optcal fers (Ths s usually referred to as the expensve last klometer [1]), roadand wreless access (BWA) has ganed much more nterest than ever. The Internatonal Telecommuncatons Unon (ITU) fgured that developng countres need to nstall more than a mllon new telephone lnes n a couple of years, and that the telecommuncatons operators expect to replace the wred nfrastructure wth wreless technology [1]. Also, there appeared more demands on gettng an access to the Internet wth personal dgtal assstant (PDA) or mole phones n pulc areas such as arports and tran statons. Ths tendency nvokes the necessty of BWA tremendously ecause deployng an nfrastructure of cale modem or hyrd fer cale n pulc areas would e costly. Based on these demands, BWA wll encompass varous servces such as voce,

15 2 vdeo on demand, telephony, and Internet access. In order for BWA to make nroads nto the market successfully for these servces, a numer of ssues, ncludng relale non-lneof-sght (NLOS) operaton, spectrum effcency, and network scalalty, should e resolved frst [2]. In ths dssertaton, several approaches are consdered to resolve these ssues. Fgure 1-1 llustrates an example of BWA. Fgure 1-1: An example of roadand wreless access (BWA). Frstly, relale NLOS operaton hghly depends on the frequency and. Mllmeter-wave and ganed great attracton n early days due to the fact that FCC allows more than 1 GHz and n the mllmeter-wave spectrum for LMDS (Local Multpont Dstruton Servces) applcatons. However, mllmeter wave channels are very vulnerale when sgnal power attenuaton ncreases and outage proalty ecomes very hgh, especally when LOS s lost. Hence, LMDS s known to e sutale only for very hgh uldngs where drect LOS lnks are attanale. Ths characterstc proaly encourages estalshng mllmeter wave lnks etween a central offce and ase statons

16 3 whle uldng UNII and wave lnks etween ase statons and suscrers. Currently, ecause of the severe power attenuaton characterstcs of mllmeter-wave channels, many telecommuncatons companes have lost nterest n mllmeter-wave and and have egun to favor UNII (Unlcensed Natonal Informaton Infrastructure) ands or lower ands. Usng UNII ands, the folage penetraton and NLOS (Non-Lne of Sght) operaton ecomes feasle wth suffcently large avalale andwdth of 300MHz. For a NLOS soluton, many degrees of freedom have een exploted. IEEE a has een the leadng standard group for maxmzng the use of degrees of freedom [3]. Fgure 1-2 shows some of the degrees of freedom to e exploted. Among these, we choose to explot modulaton and tme for the NLOS soluton n our systems; we focus on adaptve modulaton and automatc repeat request (ARQ), whch s also known as tme dversty, snce t s part of Internet Protocol (IP) standard technque. Fgure 1-2: Communcaton degrees of freedom

17 4 Adaptve technques, such as adaptve modulaton, have een extensvely studed n order to oost the spectral effcency n fadng channels. Conventonally, wreless communcaton systems wth fadng channels have een desgned consderng the worst channel stuaton to guarantee a defned mnmum acceptale performance. However, desgnng a communcaton system consderng the worst channel condton allows us to use only a small porton of the entre capacty of the channel and produces a low spectral effcency. In order to prevent ths, adaptve technologes have een ntroduced and adopted n communcaton system desgns. Although adaptve technques can lead us to very hgh spectral effcency, they ddn t draw much attenton n ther early days ecause of hardware constrants and a lack of avalalty of good channel estmaton technques. Today, these challenges are no longer dffcult prolems ecause technologes have matured. Thus, adaptve technques have een ganng more and more nterest n ndustral and academc socetes [5]. We also examne heterogeneous traffc due to multmeda servces n roadand wreless access. As roadand wreless networks encompass varous servces, such as, world-wde-we (www), voce, vdeo and data, the resultant network traffc ecomes very dynamc. For conventonal voce only transmssons, andwdths of uplnk and downlnk channels have een consdered symmetrc; now, those andwdths are unalanced and the rato also vares dynamcally ecause multmeda servces requre much larger andwdths for downlnk transmssons. To provde the hghest transport effcency n roadand networks, tme dvson duplex (TDD) s preferred to frequency dvson duplex (FDD) ecause t enales real-tme adaptaton of uplnk and downlnk andwdth accordng to the dynamc traffc pattern [26][27][42]. We call ths TDD

18 5 scheme wth dynamc oundary dynamc TDD (D-TDD) scheme, and we call the conventonal TDD scheme wth fxed oundary fxed TDD (F-TDD). Jeong et al. [43] proposed an optmal tme slot allocaton (TSA) method assumng the oundary of TDD s set once a day or once n a long tme perod. The TSA method showed etter utlzaton for unalanced traffc ut Jeong et al. dd not address how the TSA scheme would work wth dynamcally varyng traffc pattern, especally when the traffc characterstcs are dfferent cell y cell. Adjustng the uplnk and downlnk volume swtchng pont n each cell may rng hgher spectral effcency for the entre network, ut t also creates another prolem that should e resolved. By havng dfferent swtchng ponts for uplnk/downlnk n dfferent co-channel cells, D-TDD schemes are serously lmted y co-channel nterference (CCI) from ase statons n other co-channel cells. Ths s due to the fact that the propagaton from a ase staton to a ase staton suffers less attenuaton than that from suscrer antenna to a ase staton ecause the antenna heght at a ase staton s much hgher than that of a suscrer antenna. Jeong et al. [27] ntroduced a tmeslot allocaton algorthm mnmzng the outage performance wthout consderng rate and power allocaton. Whle the analyss s ased on perfect knowledge of other ase statons frame resource nformaton such as the swtchng pont and actve sector ndex, t mght e dffcult to otan perfect frame nformaton of all co-channel cells n a practcal stuaton ecause every cell wll allocate tme slots for etter spectrum effcency after t roadcasts frame nformaton. Hence, we developed a tme slot allocaton algorthm that does not need perfect frame resource nformaton ut rather oundary nformaton.

19 6 Fnally, we propose an adaptve resource allocaton algorthm for roadand wreless access. The proposed resource allocaton scheme oosts the spectral effcency of TDD/TDMA systems y adaptng to the varyng channels wth adaptve modulaton and y adaptng the uplnk/downlnk traffc alance wth D-TDD scheme. We examned the performance of our system n four cases: sngle cluster model wth D-TDD and fxed modulaton, sngle cluster model wth D-TDD and adaptve modulaton, mult-cluster model wth D-TDD and fxed modulaton, and mult-cluster model wth D-TDD and adaptve modulaton. For all these cases, we compare performance results wth F-TDD cases. Below, we ntroduce the overvew of all the adaptve technques we consdered n ths dssertaton. 1.2 Overvew of Adaptve Technques Adaptve Modulaton and Codng In the 1960 s and 1970 s, adaptve rate transmsson usng adaptve modulaton and adaptve codng, whch was ased on the assumptons of accurate channel estmaton and a relale feedack path, were frst proposed [7]. After then, Cavers proposed an adaptve modulaton scheme wth delayed feedack channel [8], and Mandelaum proposed an adaptve code rate transmsson scheme [9]. Although the adaptve modulaton technque could lead us to a very hgh effcency, the nterest n adaptve modulaton wasn t fully apprecated ecause of hardware constrants and a lack of good

20 7 channel estmaton technques. Ths made adaptve codng comned wth Hyrd automatc repeat request technques (Hyrd ARQ) the manstay of adaptve rate systems [10]. Snce there has een much progress n hardware mplementaton, adaptve modulaton technque has revved. Flp et al. proposed optmum utlzaton of channel capacty usng adaptve modulaton [11], and later they reported on the feaslty of mplementng the adaptve modulaton scheme, whch led to the wde acceptance of the adaptve modulaton technques [12]. Goldsmth et al. proposed adaptng power level, as well as modulaton scheme, to the varyng channel condton. They compared the spectral effcency of ther adaptve method to the upper ound on spectral effcency and to the spectral effcency of non-adaptve modulaton scheme. They also deal wth the transmtter s frequency when changng ts constellaton and power accordng to the channel Doppler frequency [5]. Aloun appled the adaptve modulaton technque to the Nakagam fadng channel [13], and he also appled t to smultaneous voce and data transmsson over fadng channels [14]. Ue et al. proposed symol rate and modulaton level-controlled adaptve modulaton/tdma/tdd system for hgh-t-rate wreless data transmsson [15]. Recently, Chung et al. reported aout the degrees of freedom n adaptve modulaton. They found a general form for power, BER, and data rate adaptaton that maxmzes spectral effcency [16]. Whle the development of adaptve modulaton technque was delayed due to some dffcultes, adaptve codng has een studed successvely as an adaptve rate technque. Goldsmth et al. appled adaptve codng scheme to adaptve modulaton scheme. In ther report, they clamed that trells and lattce codes desgned for addtve

21 8 whte Gaussan nose (AWGN) channels can e comned wth adaptve modulaton for fadng channels, wth the same approxmate codng gans [17]. Snce deletng the output ts of an encoder (puncturng) perodcally can change the code rate easly, rate compatle punctured convolutonal (RCPC) codes and rate compatle punctured turo (RCPT) codes have een ntroduced as adaptve codng technques [18][19]. Can et al. frst ntroduced the punctured convolutonal codes [20]. The purpose was to fnd hgh rate codes wth error correctng capaltes of good convolutonal codes ncludng large memory. These are now consdered the most convenent way of mplementng the varale rate system [21] Adaptve ARQ Schemes Recent studes on ARQ schemes adopt adaptve rate transmssons to ensure a hgher proalty n the acceptance of retransmtted data [22]. Current ssues wth regards to maxmzng throughput motvate usng adaptve rate transmssons wth ARQ technques for a hgher spectral effcency n addton to provdng a more relale transmsson [23]. Hence, data rate for each lock transmsson s decded y the estmated channel condton rather than y whether the lock corresponds to an orgnal transmsson or a retransmsson. Three asc ARQ schemes, such as stop-and-wat ARQ, go-ack-n ARQ, and selectve-repeat ARQ, are well known and have een analyzed ased on fxed rate transmsson for addtve whte Gaussan nose (AWGN) channels n [10]. Najoh et al. ntroduced ARQ technques as effectvely comned wth adaptve modulaton/tdma/ TDD to acheve hgher throughputs. They also provded some

22 9 computer smulaton results n [23]. Recently, Yun et al. provded expressons for throughput of the three ARQ technques n adaptve modulaton systems over a fadng channel, assumng lock transmssons are ndependent and dentcally dstruted [24][25]. However, analytcal approaches for the performance estmaton of adaptve rate systems comned wth ARQ schemes have not een well addressed yet Dynamc TDD Systems Frequency-dvson duplex (FDD) has een mplemented n conventonal wreless telephone systems such as GSM, IS-136, and IS-95. FDD requres a pared channel for communcaton, one for downlnk and one for uplnk. Snce the andwdths for these two channels are usually fxed, FDD has een wdely used for largely symmetrc and predctale traffc, such as voce traffc. One of the most mportant advantages n FDD reles on the fact that there s no nterference etween uplnk and downlnk sgnals when the uplnk channel and downlnk channel are suffcently separated n frequency, whch, however, ntroduces dffculty n RF hardware mplementaton n congested parts of the rado spectrum. Tme-dvson duplex (TDD), on the other hand, s eng used for asymmetrc and unpredctale pont-to-multpont dgtal data networks, such as roadand multmeda networks. TDD can handle uplnk and downlnk asymmetrcally y allocatng tme spent on uplnks and downlnks, requrng only one frequency channel for full duplex communcatons.

23 10 Recently, TDD has ganed wde attenton n fxed roadand wreless access systems ecause of several enefts: recprocty of the channel for the sngle carrer frequency used on oth uplnk and downlnk wll offer channel state nformaton (CSI) easly, RF desgn s smpler, the allocaton algorthm s more flexle for dynamc up/down traffc, and oth uplnk and downlnk use only one RF carrer [26]. TDD, agan, can e dvded nto two types: Fxed TDD (F-TDD) and Dynamc TDD (D-TDD). In the F-TDD case, tme slots allocated for uplnks and downlnks are fxed. However, D-TDD has an adjustale oundary etween the uplnk and downlnk duty cycles ased on the servce requrement. The man prolem of ths D-TDD system s the possle co-channel nterference from the downlnk sgnals of other ase statons durng the uplnk duty cycle of the desred cell. L et al. have shown that the strong cochannel nterference (CCI) can e suppressed y usng an adaptve array antenna [26]. Meanwhle, Jeong et al. have provded a smple tme slot allocaton algorthm comned wth a sectored antenna layout to suppress the strong CCI. Both schemes show promsng results on SIR outage performance [27]. 1.3 Organzaton of the Dssertaton The remander of ths dssertaton s organzed as follows. Chapter 2 deals wth a fnte-state Markov channel model and two contnuous fadng channel smulators, Jakes model and Clarke s model, for the purpose of analyses and smulatons. Chapter 3 ntroduces lock error models, the conventonal two-state error model and a mult-state error model. We descre the mult-state error model, whch s useful for the performance

24 11 analyss of adaptve modulaton systems. In Chapter 4, adaptve modulaton system comned wth automatc repeat request (ARQ) schemes s llustrated, and ts performance evaluaton s followed. In Chapter 5, dynamc tme dvson duplex (D- TDD) s ntroduced, and a co-operatve tme slot allocaton (C-TSA) s proposed wth ts performance analyss. In Chapter 6, we propose an adaptve resource allocaton algorthm for heterogeneous traffc. Fnally, Chapter 7 summarzes the man contrutons of ths dssertaton and lsts some future research drectons.

25 12 Chapter 2 FINITE STATE MARKOV CHANNEL MODEL 2.1 Introducton Whle wreless channels have een known to e much less relale than wred channels, recent research shows that fadng channels have a much larger capacty than antcpated wth tradtonal approaches [6]. Ths modern vew on fadng channels encouraged us to characterze these channels more precsely for etter dentfcaton and utlzaton of wreless channel capacty. In general, desgns communcaton systems have carred out analyses and smulaton ased on the assumpton that the sgnal power level of symols or locks at the recever are ndependent and dentcally dstruted (..d). Therefore, tradtonally t has een attempted to elmnate the channel memory usng nterleavng to support the..d assumpton, whch results n a loss of enefts assocated wth channel memory. However, n practcal stuatons, consecutve lock transmssons have more chances to e correlated, especally over a slow fadng channel. Ths correlaton should not e neglected. Therefore, the prolem has ganed much nterest, recently. Snce the Markov model s a natural way to approxmate a channel wth memory, many people have

26 13 consdered fnte-state frst-order Markov modelng for descrng a wreless communcaton channel [31-36]. In ths chapter, we frst ntroduce the relatonshp etween a physcal fadng channel and the correspondng fnte-state Markov model, whch can e used for performance evaluaton n a packet transmssons network. After we estalsh the fnte state Markov model, we also ntroduce fadng channel smulators, known as Jakes model and Clarke s model. Wth slow t-level smulatons wth Jakes model and Clarke s model, we show that the fnte-state Markov model s accurate enough to evaluate the performance of packet transmsson systems. 2.2 Fnte-State Markov Channel (FSMC) Models It s well known that the effect of fadng s multplcatve n representng aseand sgnal ~ ~ s ( t) = u~ ( t) f ( t ) 2.1 where ~ s ( t ), ~ ~ u ( t), f ( t) are the complex envelopes of channel output, channel nput, and ~ ~ ~ fadng, respectvely. And, the fadng process can e represented as, f ( t) = f ( t) jf ( t), C + ~ ~ where the n-phase and quadrature fadng processes fc ( t ), ( t ) are Gaussan, statonary, and ndependent wth the same power spectrum gven y [28] f S S

27 14 2 σ 1 for f < f m, 2 2 πf m S ( f ) = f 1 f m 0 otherwse, 2.2 where 2 σ s process power, and f m s the maxmum Doppler frequency. The correspondng autocorrelaton functon s 2 R( τ) = σ J ( 2πf τ), m where J 0( ) s the zeroth-order Bessel functon of the frst knd. Assumng the n-phase and quadrature processes are zero mean, the resultng ~ fadng process can e descred as Raylegh. The real envelope f ( t) = f ( t) follows the Raylegh dstruton Hence, the receved nstantaneous SNR γ n the Raylegh fadng channel has an exponental dstruton wth a proalty dstruton functon where γ0 0s the average receved SNR. 2 r r f = f ( r) exp 2 σ 2.4 2σ 2 1 γ f γ ( γ) = exp, γ 0, 2.5 γ0 γ0 Denotng γ (t) as receved SNR values at tme t, we can otan sampled process of receved SNR wth a fxed perod T. Generally, T represents a tme nterval sgnfcant for the communcaton system transmttng over the channel; t could e a t, a symol or a frame duraton. Then, the process γ (t) s quantzed to N levels wth respect to the set of

28 15 thresholds Γ, {0,.., N}, Γ 0 = 0 and Γ N =. Then, the resultng process γ = Γ γ ( nt ) Γ }, { 1,.., N}, s approxmately Markovan and ths s n { 1 F presented n Fgure 2-1. For parttonng SNR levels,.e., fndng threshold values Γ, {0,.., N} to complete the FSMC model, several methods have een used. Goldsmth et al. partton SNR levels ased on the BER constrants havng dfferent average fade state duratons (AFSD) [5]. Wang et al. use the equal-proalty method [31], and Zhang et al. partton SNR levels makng the AFSD equally lkely [32]. In our work, we use Zhang s equal duraton method for fndng threshold values. Fgure 2-1: The fnte-state Markov channel model Wth the maxmum Doppler frequency, f m, the level crossng rate of level, Γ,.e., the average numer of crossngs of a gven sgnal level, Γ, per unt nterval of tme n the postve (or negatve) drecton can only e otaned as n [5, 31-33] 2πΓ = Γ N( Γ ) f exp, m 2.6 γ0 γ0 where γ0 s the average SNR.

29 16 The average duraton of tme that recever SNR stays etween Γ, Γ ) s the [ 1 rato of the total tme the recever SNR remans etween, Γ ) and the total numer of such sgnal segments. It can e expressed as [32] [ Γ 1 π τ =, 2.7 N Γ ) + N( Γ ) ( 1 where π s the steady-state proalty of each state that can e otaned as Γ Γ 1 γ π = f γ ( γ) dγ = exp( ) dγ, 2.8 Γ 1 Γ 1 γ γ 0 0 where = 1,..., N. In order to make the average tme duraton of each state large enough to cover the frame tme, T F, we can set τ = ct wth c (>1) a constant. Ths way, we can fnd all F threshold values Γ, {0,.., N} ensurng that the SNR range of each state s equally large enough to cover the frame duraton. In our work, we use a constant c etween 3 and 8, as recommended n [32]. Assumng channel state transton occurs etween adjacent states only n slow fadng channels, we can fnd transton proaltes, t, j = Pr{ Sn+ 1 = j Sn = } as [5, 31-33]

30 t t t t, +,, j, N( Γ + 1) TF 1, π N( Γ ) TF 1, π 0, = 1 N t, j j= 1, j, = 1,..., N 1, = 2,..., N j 2, = 1,..., N Hence, the transton matrx T can e completed as: T t1,1 t2,1 0 = t t t 1,2 2,2 3, t t t 0 2,3 3,3 3,4 0 0 t t 0 0 3,4 3,4 0 0 t t N 1, N N, N 1 t N 1, N t N, N As we see from these equatons, elements of the transton matrx T are proportonal to the fade rate f T m F. In [33], Bach et al. stated that the resultng matrx T s acceptale only f the fade rate s very low, and also addressed that we would need more complex methods for hgh fade rate. As we focus on slow fadng channel, we use these results n our analyss. 2.3 Smulaton of Fadng Channel Models In ths secton, we llustrate wdely used technques of generatng a fadng envelope for smulaton purpose. One way of generatng a fadng envelope s to construct a fadng envelope from n-phase and quadrature Gaussan nose sources [29] ecause the

31 18 envelope of a complex Gaussan nose process has a Raylegh proalty densty functon (PDF). In ths technque, the Doppler spectrum s provded y applyng approprate low-pass flters to the Gaussan nose sources. Ths technque s known as Clarke s model. Smth [30] demonstrated a methodology to mplement Clarke s model. An alternatve way to modelng wth fltered complex Gaussan nose s to approxmate the Raylegh fadng process y summng up a numer of complex snusods. Ths s known as Jakes model ecause ths was proposed y Wllam Jakes for the smulaton of fadng mole rado channels [28] Clarke s Model As n Fgure 2-2, two ndependent Gaussan low pass nose sources are used to generate n-phase and quadrature components of a fadng envelope. By usng the Doppler Flter, tme doman waveforms of Doppler fadng can e generated y usng an Inverse Fast Fourer transform (IFFT) at the fnal stage of the smulator.

32 19 Fgure 2-2: Clarke s fadng smulator usng quadrature ampltude modulaton wth aseand Doppler flter The detaled procedure to mplement the smulator s ntroduced n [30]. Fgure 2-3 shows exemplary fadng envelopes generated y the Clarke s model. We generated two envelopes wth dfferent speeds of vehcles,.e., dfferent maxmum Doppler frequences.

33 20 Fgure 2-3: Fadng envelope levels generated y Clarke s model for dfferent vehcle speeds wth f c =1.9GHz: (a) v = 5km/h (f D = 8.796), () v = 20km/h (f D = ) Jakes Model Early n the 1970 s, Jakes suggested a very effectve channel smulator ased on the sums of snusods [28]. Ths method assumes equal strength multpath components. Fgure 2-4 shows Jakes smulator, whch produces a fadng envelope y summng a numer of low frequency oscllators. Choosng phases α = 0 and β = πn / M wll make < g 2 Q ( t) >= M < g 2 I ( t) >= M + 1, and < g ( t) g ( t) >= 0. The detaled procedure to I Q mplement ths smulator s ntroduced n [28].

34 21 Fgure 2-4: Jakes fadng smulator wth a sum of snusods We also plotted two dfferent fadng envelopes generated y ths Jakes model wth dfferent vehcle speeds of 5km/h and 20km/h for carrer frequency, f c = 1.9GHz, n Fgure 2-5.

35 22 Fgure 2-5: Fadng envelope levels generated y Jakes model for dfferent vehcle speeds wth f c =1.9GHz: (a) v = 5km/h (f D = 8.796), () v = 20km/h (f D = ) 2.4 Numercal Results In ths secton, we present some numercal results for the Markov channel model wth dfferent condtons. We consder the length of frame, N F = 424 symols, whch ncludes preamle N C = 40 symols, and packet tran, N P = 384 symols. For smulatons, we generate a fadng process ased on oth Clarke s and Jakes models [28][29][30] for several fade rates at vehcular speeds 5km/h and 20km/h wth 100k/s and 1M/s transmsson rates at a carrer frequency of 1.9 GHz. We made a fadng envelope for symols and, hence, we could smulate the fadng channel for

36 aout frames. The normalzed autocorrelaton functons of generated envelopes y 23 oth Jakes model and Clarke s model are plotted aganst the normalzed tme delay f m τ n Fgure 2-6. Fgure 2-6: Autocorrelaton of the real and magnary components of the receved complex envelope for Jakes smulator and Clarke's smulator Fgure 2-7 and Fgure 2-8 show Markov channel model parameters, π and, j = 1,..., N, otaned y smulaton and analytcal model. In order to otan the smulaton results, we generate fadng envelopes wth Clarke s model and count level crossngs to fnd all the parameters. In Fgure 2-7, the vehcular speed s set to 5km/h t,, j

37 wth transmsson rate of 100k/s. The fade rate s f mtf 24 = We found threshold values settng the average tme duraton of each state, τ = 3T F. In Fgure 2-8, the vehcular speed s set to 5km/h wth transmsson rate of 1M/s. The fade rate s f mtf = Snce the fade rate s very slow n ths case, a very large numer of states (N=53) s otaned even though we set τ = 8T F. In oth fgures, we fnd the analytcal model and smulaton results present a good agreement. Fgure 2-7: Markov channel model; v=5km/h, R=100k/s, N F =424 symols, τ =3T F, f m T F =0.0373, N=11; (o) smulaton, (x) analytcal model

38 25 Fgure 2-8: Markov channel model; v=5km/h, R=1M/s, N F =424 symols, τ =8T F, f m T F =0.0037, N=53; (o) smulaton, (x) analytcal model 2.5 Conclusons Ths chapter ntroduced a relatonshp etween a physcal fadng channel model wth fnte-state Markov model for packet transmsson networks. We used Zhang s partton method wth a crteron ased on the average duraton of each state [32]. Ths fnte state Markov model s useful n performance analyss avodng tme-consumng tlevel smulatons. We also ntroduced channel smulators, known as Jakes model and Clarke s model for t-level smulatons; we use these t-level smulators to show our Markov model approxmates the physcal channel accurate enough for performance evaluatons.

39 Chapter 3 MARKOV ERROR MODEL FOR BLOCK TRANSMISSIONS IN ADAPTIVE MODULATION SYSTEMS 3.1 Introducton In the prevous chapter, we studed the relatonshp etween fadng channel and fnte-state Markov model. The fadng level, whch can e represented y the recever SNR, s parttoned ased on the average fade state duraton. However, qute often the value of the channel envelope s not of drect nterest, ut rather some nonlnear functon of t, whch depends on the transmsson technques, modulaton, codng, and so forth [36]. For the purpose of throughput analyss, the channel model has een smplfed y the Glert-Ellot, whch conssts of ad and good states only [37][38]. It s known that lock error process s a good approxmaton for the purpose of performance evaluaton over suffcently slow fadng channels [36]. However, ths extremely smple model s typcally not realstc n that t does not explan the dynamc change of spectral effcency when adaptve modulaton s deployed.

40 27 In ths chapter, we nvestgate the ehavor of lock transmssons n adaptve modulaton systems on fadng channels. Accommodatng adaptve modulaton systems, we may experence dfferent spectral effcency for each lock transmsson ased on the modulaton format used for the correspondng lock. Hence, nstead of the conventonal two-state Glert-Ellot model, we splt a good state nto multple good states of dfferent spectral effcency [ts/s/hz]. Some smulaton results show that our lock error model n adaptve modulaton systems s a good approxmaton for slow fadng channels. 3.2 Error Models for Fadng Channels In ths secton, we nvestgate the lock error process whch descres the success/falure of lock data transmssons. We show that the conventonal two-state error model, whch has een used n fxed rate systems, can e extended to a mult-state error model n adaptve modulaton systems. If we denote α = α(t ) as a dscrete verson of fadng process wth samplng tme of T seconds, we have very small fade rate ( f m T << 1, where f m s the maxmum Doppler frequency and T s the samplng tme) for a moderately hgh data rate and general envronments (e.g., 5 f m T = for data rate of 100kps, carrer frequency, f C of 2GHz and vehcular speed of 5km/h). Hence, we may assume that the channel s constant durng a symol nterval. However, for the duraton of a data lock,

41 whch s comprsed of several hundreds or thousands of symols, ths mght not always e true. 28 Fgure 3-1: Glert-Ellot model: G = good state, B = ad state Two-state Error Model We denote β as a nary sequence wth value of 0 f a lock s error free and 1 f a lock contans error symol. Fgure 3-1 shows the Glert-Ellot model for lock error process. Zorz et al. [35] consdered a smple threshold model where s the power threshold, F=1/ s the fadng margn, and v s the value of fadng envelope of the frst symol n a lock. 2 0, f v >, β = , f v, Ths threshold model s consdered to e oversmplfed n the case when the fadng envelope vares sgnfcantly durng lock tme. Hence, Zorz et al. took nto

42 account all the detals of the modulaton format n [36], defnng ς as a nary process wth values of 1 f symol j s n error and 0, otherwse where P e s the condtonal symol error proalty for value of envelope, v j, whch s also dependent on modulaton format. Assumng that we use (N,k) lock code of t-ts error correcton capalty, the lock error process can e wrtten as 1, wth proalty Pe ( v j ), ς = 3.2 0, wth proalty 1- Pe ( v j ), where we gnored undetected errors for smplcty. N 1, f > = ς t, j= ( 1) N + 1 j β 3.3 0, otherwse, For adaptve modulaton systems, we can change modulaton format accordng to the channel state. Assumng we have multple transmsson modes wth a set, R = {R 1, R 2,, R L }, we can change the transmsson modulaton format accordng to the estmated channel state satsfyng the requred BER or FER performance. In ths way, we can have dfferent spectral effcency for each transmsson ased on the channel status. However aove Eq. 3.3 provdes only the nformaton aout success or fal and t does not provde the nformaton of spectral effcency n the adaptve modulaton systems. Hence, we need to modfyς and β, properly for adaptve modulaton systems. The detal of the adaptve modulaton system wll e descred n Chapter 4. 29

43 3.2.2 Mult-State Error Model 30 In 1984, Metzner [47] ntroduced a mult-state model n order to explan varyng channel condtons usng one good state and N ad states. The N ad states dffer n that they have dfferent duratons wth dfferent proaltes of returnng to the good state. Our mult-state model s dfferent from hs model n that we have one ad state and L good states. The method of dfferng the L good states s also dfferent from hs model. The L good states dffer n that they have dfferent SNR levels and hence have dfferent spectral effcency n adaptve modulaton systems. Though we may comne these two models to uld L good states and N ad states Markov model, we leave t as a future research topc. Fgure 3-2 shows the extended Markov error model for adaptve modulaton systems; we call our model an extended verson of lock process ecause the good state of the two-state model s extended to L good states n our model. Fgure 3-2: Extended Markov error model

44 31 Assumng a frame can contan one BPSK-modulated packet n t, we can say that a frame can contan M ( k) = k, k {1,.., L} packets n t when we use k-th modulaton format n the set: R = {R 1, R 2,, R L } = {BPSK, QPSK, 8-QAM, 16-QAM,, 256- QAM}. Denotng F n as the numer of successfully transmtted packets for each frame, a dscrete Markov error model for lock falure/l-success process, F n, can e descred as F n 0, = M ( ϕ ), wth proalty P F wth proalty 1- P, F, 3.4 where ϕ ={k R k s assgned for -th state}, k { 1,, L}, and F P s the frame error rate (FER) wth -th modulaton format, = 1, 2,, L. The FER n adaptve modulaton systems wll e dealt wth n Chapter 4 n detal. Representng B as a state of falure, and G k, k { 1,, L} as a state when modulaton format R k s selected and M(k) packets are successfully transmtted, we can express, P, the transton proalty matrx of the aove falure/l-success process as pbb pbg pbg pbg pbg p L 1 BGL pg B pg G pg G p G B pg G pg G pg G PBB PBG P = = pg B 0 pg G pg G PGB PGG pg L B pgl G p 1 1 L 1 GLGL pg L B pgl G p 1 L 1 GLGL

45 32 Snce we assume the channel fadng process transts only to adjacent state, the transton etween error free states occurs only etween adjacent states. In order to complete ths error model, we need to fnd elements of the state transton proalty matrx as elow:, n} Pr{error at t n} 1,error at t n Pr{error at t n} error at t 1 n Pr{error at t 1 1 1, 1 2, , ,1 1 1 = = = = = = = + = = = + = = N F N N j F j j F k F k k F F F F F F BB P P t P P P t P P t P P t P p π π π π π π 3.6, 1 1, = = = = N F N k F j j F BG P P t P p j k π π ϕ,l,, k = 2 1, 3.7, 1, = = = = k F k N j F j j F B G k P P t P p ϕ ϕ π π,l,, k = 2 1, 3.8

46 33 p GkGl = ϕ= kϕ j= l π P ϕ = k F, j π P t F P F j, k, l = 1, 2,,L. 3.9 Wth k p = p P and p ( k) = 1, we can fnd the steady state proalty, p = [ pb, pg, p 1 GL modulaton systems. ], whch completes the extended Markov error model n adaptve 3.3 Numercal Results We compare the steady state proalty and transton proaltes of our Markov error model wth those otaned from Monte-Carlo smulatons. We have multple transmsson modes wth a set, R = {BPSK, QPSK, 8-QAM, 16-QAM, 32- QAM, 64-QAM, 128-QAM, 256QAM}. We change the transmsson modulaton format accordng to the estmated channel state, whch corresponds to recever SNR per symol, satsfyng the requred FER of 0.1. For comparson, we use 424 symols for the frame length; 40 symols for preamle, and 384 symols for packet tran. We set the average tme duraton for each packet as τ =3T F to satsfy lock fadng assumpton; the numer of states for channel model s decded y ths value. For the same vehcle speed of 5km/h, we change the data rate as 100k/s and 1M/s to have dfferent fade rates, f m T F, and , respectvely.

47 34 Fgure 3-3 and Fgure 3-4 show Markov error model parameters, steady state proaltes and state transton proaltes, otaned y smulaton and the analytcal model. The smulaton settng s the same as n the Markov channel model smulaton. In order to otan the smulaton result to compare wth the analytcal results, we check errors n symol level and declare that the frame s lost f the frame has more errors than the numer of correctale errors defned y the lock code. After that, we counted all the transtons to calculate the transton proaltes; calculatng the steady state proalty s straghtforward. Comparng Fgure 3-3 and Fgure 3-4, we oserve the error model gves a etter agreement when the fade rate s very slow. Ths results show that our Markov error model would e useful n the performance evaluaton of communcaton systems n slow fadng channels.

48 Fgure 3-3: Markov error model; v=5km/h, R=100k/s, N F =424 symols, τ =3T F, f m T F =0.0373; (o) smulaton, (x) analytcal model 35

49 Fgure 3-4: Markov error model; v=5km/h, R=1M/s, N F =424 symols, τ =3T F, f m T F =0.0037; (o) smulaton, (x) analytcal model 36

50 3.4 Conclusons 37 In ths chapter, we nvestgated adaptve modulaton systems and the correspondng error model n correlated slow fadng channels. We provded a mult-state error model comprsed of one ad state and multple good states wth dfferent spectral effcency levels. We used the equal-duraton partton method to model the correlated fadng channel, whch not only reduces the quantzaton error on each state, ut also guarantees the lock fadng channel assumpton. Investgatng our model and smulatons results, we showed that our Markov error model s sutale for descrng the lock error process n adaptve modulaton systems over a correlated slow fadng channel though use of ths model may not e so accurate over fast fadng channels. Ths Markov model wll e useful n the system desgn for slow fadng channels. Especally, ths model can e used for uffer analyss that uses adaptve modulaton systems. Wth or wthout ARQ scheme, we can smply estmate the throughput adaptve modulaton systems n slow fadng channels wth our Markov model. The estmated throughput performance can e used as effectve capacty or servce rate for the uffer analyss.

51 Chapter 4 ADAPTIVE RATE TRANSMISSIONS COMBINED WITH ARQ 4.1 Introducton In accordance wth the demand for roust and spectrally effcent communcatons systems desgn for fadng channels, adaptve modulaton/codng and automatc repeat request (ARQ) have een deployed n HIPERLAN/2, IEEE a, and IEEE a [4][3]. Adaptve modulaton systems, whch were not fully apprecated n ther early days, have een revved wth advances n hardware and channel estmaton technques. Whle adaptve codng was the manstay of adaptve rate systems n the early days, several parameters, such as, modulaton format, codng, and power are all varales for hgher spectral effcency n adaptve systems today [5][17][40]. Although adaptve technques can reduce the percentage of total outage, sgnals are stll error-prone n wreless channels ecause of the vulnerale nature of a wreless medum. Therefore, forward error correcton (FEC) and automatc repeat request (ARQ) technques have een wdely used for relale communcatons systems. Recent studes on ARQ schemes adopt adaptve rate transmssons to ensure a hgher proalty n the acceptance of retransmtted data [22]. Current ssues wth regard to maxmzng throughput motvate usng adaptve rate transmssons wth ARQ technques for a hgher

52 39 spectral effcency n addton to provdng a more relale transmsson [23]. Three asc ARQ schemes, such as, stop-and-wat ARQ, go-ack-n ARQ, and selectve-repeat ARQ, are well known and have een analyzed ased on fxed rate transmsson for addtve whte Gaussan nose (AWGN) channels n [10]. Najoh et al. ntroduced ARQ technques as effectvely comned wth adaptve modulaton/tdma/ TDD to acheve hgher throughputs. They also provded some computer smulaton results n [23]. Recently, Yun et al. provded expressons for throughput of three ARQ technques n adaptve modulaton systems over a fadng channel, assumng lock transmssons are ndependent and dentcally dstruted [24][25]. However, analytcal approaches for the performance estmaton of adaptve rate systems comned wth ARQ schemes have not een well addressed as yet. 4.2 System Confguratons Fgure 4-1 shows the system confguraton of ARQ technque comned wth adaptve modulaton. At the recever, the sgnal level s measured, and ths nformaton s sent to the transmtter va a feedack channel. Based on the channel state nformaton, the modulaton level s selected frst, and then accordngly the numer of packets per frame s decded upon. Goldsmth et al. [5] addressed that BER remans at the target level f the total delay of the channel estmator and feedack channel s less than 0.001λ / v to 0.01λ / v for the target BER values of 10-6 and Here, v s the vehcle speed and λ s the sgnal wavelength. Snce we assume lnk length n our computer smulatons and analyses s less than 3 km and vehcle speed s less than 20km/h usng

53 GHz and carrer sgnals, 0.001λ / v s aout 28.4 µsec, whch s larger than the feedack delay, 10 µsec, n our smulatons. Hereafter, we assume the transmtter always has perfect channel state nformaton. Fgure 4-1: Confguraton of adaptve modulaton systems Wth an exact t error rate (BER) equaton for a two-dmensonal ampltude modulaton at a certan receved SNR, γ, n [39], we can fnd the mnmum receved SNR value for each modulaton scheme, satsfyng the target frame error rate (FER). Assumng we have L dfferent modulaton levels, we can fnd L+1 threshold values of recever SNR, A k, k {0,.., L}, A 0, and A L+1 =, wth A k as the threshold value of recever 0 = SNR, usng k-th level modulaton format ased on a target FER requrement. In our analyss and smulatons, a transmtter does not transmt packets when the SNR level s elow A 1 n order to reduce the numer of retransmssons. Havng dfferent modulaton formats for each frame, we can transmt a dfferent numer of packets, M ( k), k {1,.., L},

54 41 durng the same tme perod. Fgure 4-2 shows an exemplary frame structure, whch adopts pure adaptve modulaton wthout any forward error correcton (FEC) code. Each frame has N C BPSK-modulated symols for preamle, comprsng plot tone, and control ts, and N P symols of packet tran, whch conssts of multple packets accordng to the modulaton format. Makng the packet tran have the same numer of data symols as a BPSK-modulated packet, we can have M ( k) = k, k {1,.., L}, wth a modulaton set: R = {R 1, R 2,, R L } = {BPSK, QPSK, 8-QAM, 16-QAM,, 256-QAM}. Fgure 4-2: Exemplary frame structure for adaptve modulaton systems We may use FEC code along wth adaptve modulaton as n IEEE a, IEEE a, and other standards [4][3]. Havng FEC along wth adaptve modulaton wll

55 42 offer dfferent SNR threshold values for the same target FER requrement. The followng analyss wll e useful n performance evaluaton of oth adaptve modulaton systems (AMS) and adaptve modulaton systems wth FEC (AMS/FEC). For numercal examples, we consder adaptve modulaton systems wthout FEC. Havng P B k (γ) as the exact BER of k-th modulaton format ased on the exact BER calculaton n [39], we can calculate the exact packet error rate (PER), P P k (γ), as P ( γ) P γ)) P k B N = 1 (1 ( P, k 1,..., L k =, 4.1 where wrtten as N P s the length of each packet. If we assume a lock code wth t-ts error correcton capalty, PER can e P P k t NP B l B N P l ( γ) = 1 ( Pk ( γ) )( 1 Pk ( γ) ), k = 1,..., L. 4.2 l = 0 l Hence, the exact frame error rate (FER) can e otaned as F C P ( γ) k B N P M ( k ) = 1 (1 P1 ( γ)) (1 Pk ( γ)), k 1,..., L =, 4.3 where M ( k), k {1,.., L} s the numer of packets n a frame wth k-th modulaton format. As n [13][17][40] for the BER approxmaton and n [41] for the PER approxmaton, we can also fnd FER approxmaton as elow P F k 1, ( γ) = a k exp( g k γ), f 0 < γ < γ, f γ γ, pk pk k = 1,..., L, 4.4

56 43 where a, and γ are found y least-squares curve fttng as lsted n Tale 4-1. k g k pk Tale 4-1: Parameters for approxmated FER equatons; N C = 40, N P = 384 symols Modulaton Rate (ts/sym.) a g γ p (db) Mode 1 BPSK Mode 2 QPSK Mode 3 8-QAM Mode 4 16-QAM Mode 5 32-QAM Mode 6 64-QAM Mode QAM Mode QAM Fgure 4-3 shows how approxmated FER curves wth the aove parameters ft the exact FER curves. Hence, we wll use Eq. 4.4 as an approxmate FER for our analyss, nstead of usng the exact FER equaton (2) ased on the exact BER equaton n [39]. Fgure 4-3: Curve fttng test to fnd parameters: a, g, γ p as shown n Tale 4-1. (-) exact FER curves, (*) approxmated FER curves, and (--) exact BER curves for {BPSK, QPSK, 8-QAM, 16-QAM, 32-QAM, 64-QAM, 128-QAM, 256-QAM}

57 44 When we use adaptve modulaton, each state wll e assgned a specfc modulaton format that can offer the hghest spectral effcency and also satsfes the target FER. If we denoteϕ ={k R k s assgned for -th state}, = 1, 2,, N, k { 1,, L}, the average FER for each state can e otaned usng Eq. 4.4 as P F 1 = π 1 = π Γ Γ 1 Γ Γ 1 = π γ a 1 0 ϕ g a exp( g a γ ϕ ϕ ϕ 0 ϕ exp( ( g exp( ( g γ) exp( γ ϕ ) γdγ γ ϕ ) Γ γ 0 γ γ 0 ) dγ 1 ) exp( ( g ϕ 1 + ) Γ ). γ0 4.5 where γ 0 s the average receved SNR. We present the average FER for each state when the average recever SNR s 18dB and the target FER s 0.1 n Tale 4-2. Tale 4-2: Example of threshold levels and average FER; average SNR, γ 0 =18dB target FER=0.1, and τ =3T F Γ Selected Modulaton State State BPSK e-5 State QPSK e-6 State QAM State QAM State QAM e-5 State QAM State QAM State QAM State 10 ~ 64-QAM F P

58 4.3 Throughput Performance of AMS comned wth ARQ 45 As we descred n secton 4.2, we consder fxed-length packets, whch are components of frames, as retransmsson unts. We change the transmsson rate for each frame, so that we can send multple packets n one frame wth a specfc modulaton format, ased on the channel nformaton on sgnal-to-nose rato (SNR) provded y the recever end, where perfect coherent detecton s assumed. Transmttng as many packets per frame as the channel allows, we can maxmze the spectral effcency of the channel. Each erroneously receved packet for each frame wll e retransmtted mmedately when acknowledged at the transmtter. In ths way, each frame transmsson can nclude oth retransmttng current packets and new packets. Btmap-ased acknowledgement s used to show sequence numers of erroneously receved packets for each receved frame. The feedack channel for ACK/NACK s assumed free of errors, and the fadng level s also assumed to e nvarale durng a frame perod. We assume that the transmtter uffer s always loaded wth some packets for the next frame transmsson so that we can only nvestgate the effect of correlated fadng channels on the throughput performance of ARQ schemes comned wth adaptve rate transmssons. Furthermore, we do not mpose any restrctons on the numer of retrals of a lost packet. Conventonally, the throughput performance of a retransmsson request system s otaned y estmatng the average numer of tmes a packet must e transmtted efore t s accepted y the recever, T r. Assumng all packet errors are detected at the recever

59 and the packet error rate (PER) s a constant over long perod, the dervaton of T r s as follows: 46 T r = (1 PER) + 2 PER(1- PER) + + k PER = (1 PER) k = 1 k(per) k-1 1 =. 1 PER k-1 (1 PER) Based on ths value, we could otan the throughput performance of oth goack-n ARQ and selectve-repeat ARQ schemes. η k 1 =, 1 ( 1) n + Tr N GBN 4.7 η SR k 1 =. 4.8 n Tr However, ths approach s not applcale to adaptve modulaton systems over fadng channels, where transmsson rate and PER are varyng. For example, let s assume a typcal stuaton where a frame havng two packets modulated y QPSK has een lost and after a round-trp-delay, the transmtter s notfed that the frame has een lost and retransmts those two packets. If the estmated SNR s stll aove the threshold value for QPSK, the transmtter wll send those two packets n one frame agan ut, f the SNR s lower than that threshold, the transmtter wll e sendng only one packet for the current frame and one remanng packet wll e sent n the susequent frame. Ths example explans why the average numer of retransmssons alone cannot estmate throughput performance n adaptve modulaton systems. In ths chapter, we ntroduce a method of

60 estmatng the throughput performance of adaptve modulaton systems y fndng the average numer of successfully transmtted packets over a unt perod Adaptve Modulaton comned wth Go-ack-N ARQ In an adaptve modulaton system, transmtter forms a frame of multple packets ased on the estmated channel status and sends each frame to the recever contnuously, and goes ack, a round-trp delay, m frames upon recepton of a NAK (negatve acknowledgement) as n the conventonal fxed modulaton systems. When a frame s found erroneous, all the packets n the erroneous frame and all the packets of susequent N-1 frames are dscarded as n Fgure 4-4 and Fgure 4-5. Fgure 4-4: Go-ack-N ARQ example In Fgure 4-5, we present a typcal example of transmsson sequence wth m=3. Numers, 0,1,2,3,4,5 are falure/l-success process, F n. A long sequence can e dvded y several cycles, whch s started y an erroneous frame and ended y the frame just efore another erroneous one [46].

61 48 Fgure 4-5: A typcal sequence example for Go-ack-N ARQ Defnng normalzed throughput as the average numer of data packets accepted y the recever n a tme unt, we could express the normalzed throughput for go-ack-n ARQ wth adaptve modulaton as accepted packets durng one cycle η = elapsed tme for one cycle = E E F P +, 4.9 m where E F s the average accepted numer of frames n a sngle cycle and E P s the average accepted packets n a sngle cycle; the m-1 frames followng after the erroneous one are not ncluded n E F ecause they are dscarded at the recever. To fnd E P and E F, we need to consder the m-step state transton matrx P(m)=P m, wth elements p ( m) = Pr{ S + = j S = k}, k, j {1,.., }. Denotng Pˆ as a k, j n m n L matrx where the frst row s all zeros ut other elements are the same wth state transton matrx, P, we can fnd the average numer of transmtted frames n a sngle cycle, E F as E F = = ( m) B, G j j= 1 k= 1 L L p p ( m) B, G j j= 1 k= 1 k Pr{state starts at G k pˆ G j, B ( k) + 0 p B, B, j and goes ack to B after k frame} 4.10 where pˆ, ( k) s (j+1,1) element of G j B ˆ ˆ k ( ) =. P k P

62 49 Wth M ( k), k {1,.., L} as the numer of packets fllng a frame n each state, the average numer of accepted packets n a sngle cycle can e expressed as B, G j j= 1 k= 1 M ( j) ( q W ( k) ) L E P = p ( m) G j, B, 4.11 q q= 1 k where W ( ) s (j+1,1) element of k ˆ M ( 1) M ( L) W ( ) = ( P Q), Q = dag(1, q, q ). G j, B k For nfnte sums n Eq and Eq. 4.11, whch ultmately approach asymptotc values, we calculate the summaton untl the ncrement s less than 10-4 for our numercal results. Snce calculatng Eq s numercally ntensve, we may fnd the approxmate value of E P as L j= 1 M ( j) pg j E P E F. L 4.12 p j= 1 G j As a specal case of a two-state Markov error model wth M 1 =1, p P = p p p Q = dag(1, q M (1) 1 ) = 0 0 q, we could get E P = p 1,2 (m)/p 2,1 and E F = p 1,2 (m)/p 2,1 from Eq and Eq Ths result s the same as the throughput effcency formula n [23]. Moreover, f we assume the error process mechansm as a random error channel,.e., p 1,2 (m) = p and p 2,1 = 1-p where p s the proalty of a packet to e successfully receved, the throughput equaton reduces to the Ln s throughput formula of p/[p+m(1- p)] [9].

63 4.3.2 Adaptve Modulaton comned wth Selectve-repeat ARQ 50 The transmtter sends packets to the recever contnuously and retransmts only those packets that were negatvely acknowledged, as shown n Fgure 4-6. Fgure 4-6: Selectve-repeat ARQ example For selectve-repeat ARQ, packets susequent to an erroneously receved packet n a frame are not dscarded, and every correctly receved packet s kept n the recever uffer. The recever should release error-free receved packets n consecutve order from the recever uffer. Hence, suffcent recever ufferng must e made avalale. The average numer of successfully accepted packets per each frame can e smply expressed as, q L ( p ( T Q) 1) L q= 1 η = lm = M ( j) pg, L j L j= where p = p, p, p ], T s the state transton matrx, 1 = [ 1, 1, 1,, 1] and [ B G1 GL M ( 1) M (2) M ( L) Q = dag(1, q, q, q ).

64 4.4 Cross-Layer Optmzaton 51 Our analyss can e used to fnd optmal values of several parameters maxmzng the throughput performance. Snce the throughput,η, s a functon of average SNR, target FER and fade rate, f mtf n our analyss, we can fnd the optmal target FER as, F, opt P t = arg maxη ( P arg et F 0< P t arg et < 1 F t arg et ) 4.14 where F Pt arg et s the target FER and F opt P, t et arg s the optmal target FER. 4.5 Numercal Results Fgure 4-7 to Fgure 4-10 show throughput performance for dfferent vehcular speeds and dfferent fade rates. As can e oserved from the fgures, throughput performance s more dependent on fade rate, f mtf than the vehcular speed. Comparng smulaton results and analytcal results, we can see that our model s very accurate for slow fadng channel wth the fade rate f mtf = as n Fgure 4-7. The faster the fade rate gets wth f mtf =0.0149, , and , the wder gap we can fnd etween analyses and smulatons results, as expected from the autocorrelaton functon n Fgure 2-6. When f mtf = and f mtf =0.0149, the frst symol of a frame and the last symol of a frame are hghly correlated and lock fadng assumpton s acceptale. However, f the fade rate ncreases, lock fadng s not guaranteed, and the sgnal level

65 52 vares greatly durng frame tme. Smulaton results also show that the throughput performance gap etween GBN-ARQ and SR-ARQ s very small when the fade rate s very small as n Fgure 4-7. However, the gap ncreases as the fade rate ncreases as n Fgure Fgure Ths shows that SR-ARQ outperforms GBN-ARQ n the condton of hgher retransmsson proalty. Fgure 4-7: Throughput performance for v = 5km/h, R=1M/s, N F =424 symols, f m T F = Fgure 4-7 shows throughput performance of SR-ARQ and GBN-ARQ when the vehcular speed s 5km/h, data rate s 1Mps, and the fade rate s Analytcal results and smulaton results are almost matchng when the fade rate s very small.

66 53 Fgure 4-8: Throughput performance for v = 20km/h, R=1M/s, N F =424 symols, f m T F = Fgure 4-8 shows throughput performance of SR-ARQ and GBN-ARQ when the vehcular speed s 20km/h, data rate s 1Mps, and the fade rate s Wth hgher fade rate, analytcal results show etter performance than smulaton results. Ths results from the mult-state lock error process that gnores fade varatons durng frame tme.

67 54 Fgure 4-9: Throughput performance for v = 10km/h, R=100k/s, N F =424 symols, f m T F = Fgure 4-9 shows throughput performance of SR-ARQ and GBN-ARQ when the vehcular speed s 10km/h, data rate s 100kps, and the fade rate s As the fade rate goes up, the gap etween smulaton results and analytcal results ncreases.

68 55 Fgure 4-10: Throughput performance for v = 20km/h, R=100k/s, N F =424 symols, f m T F = Fgure 4-10 shows throughput performance of SR-ARQ and GBN-ARQ when the vehcular speed s 20km/h, data rate s 100kps, and the fade rate s Fgure 4-11 shows the throughput performance accordng to the target FER and average receved SNR values. For ths throughput performance, we set the vehcular speed to 5km/h and the transmsson rate to 1M/s. For γ 0 =20dB and γ 0 =25dB, the throughput performance s maxmum wth a target FER as 0.1 and for γ 0 =15dB, the throughput performance s maxmum wth a target FER as 0.5. Snce we assumed that uffers have nfnte length and there s no packet dropped at the uffer y overflow, we

69 may otan a dfferent result f we apply our model n a fnte uffer system, consderng oth dropped packets at the uffer and lost packets over the fadng channel. 56 Fgure 4-11: Fndng an optmal target FER for maxmum throughput performance 4.6 Conclusons In ths chapter, we nvestgated adaptve modulaton systems and error model n correlated slow fadng channels. We provded a mult-state error model comprsed of one ad state and multple good states wth dfferent spectral effcency levels. We used the

70 57 equal-duraton partton method to model correlated fadng channel, whch not only reduces the quantzaton error on each state, ut also guarantees the lock fadng channel assumpton. Investgatng our model and smulatons results, we showed that our Markov error model s sutale for descrng error model n a correlated slow fadng channel and that t may not e so accurate under a fast fadng condton. Based on the Markov error model, we present throughput estmaton methods of Go-ack-N ARQ and Selectve- Repeat ARQ schemes n adaptve modulaton systems (AMS). Smulaton results show that our model s sutale n estmatng throughput performance of ARQ schemes n adaptve modulaton systems n slow fadng channels. The throughput expresson for Goack-N ARQ n adaptve modulaton systems s the generalzaton of Leung s throughput equaton, whch was also a generalzaton of a random error model to a two-state Markov error model [46]. The throughput performance otaned y our Markov error model can e used as the servce rate n the uffer analyss of adaptve modulaton systems. Our performance evaluaton method wll e useful n cross-layer optmzaton.

71 58 Chapter 5 DYNAMIC TIME DIVISION DUPLEX SCHEME FOR HETEROGENEOUS TRAFFIC 5.1 Introducton As roadand wreless networks encompass varous servces, such as, worldwde-we (www), voce, vdeo and data, network traffc ecomes very dynamc. In other words, volumes of uplnk and downlnk streams, whch have een consdered symmetrc for conventonal voce transmssons, are unalanced and the rato s tme varyng. To provde the hghest transport effcency n roadand networks, tme dvson duplex (TDD) s preferred to frequency dvson duplex (FDD) ecause t enales real-tme adaptaton of uplnk and downlnk andwdth accordng to the dynamc traffc pattern [26][27][42]. Even though FDD also can e used for asymmetrc traffc, downlnk and uplnk channel andwdths should e matched to the rato of downlnk and uplnk traffc and, moreover, channel ands cannot e adjusted dynamcally n response to the varyng rato of downlnk and uplnk traffc due to hardware lmtatons. The rato of uplnk and downlnk streams s fxed for FDD. Jeong et al. [43] proposed the optmal tme slot allocaton (TSA) method assumng the oundary of TDD s set once a day or once n a long tme perod. The TSA

72 59 method showed etter utlzaton for unalanced traffc, ut they dd not address how the TSA scheme would work wth a dynamcally varyng traffc pattern, especally when the traffc characterstcs are dfferent cell y cell. For roadand networks, the traffc pattern n each cell s dfferent ased on the type of servces that suscrers are recevng n each cell. Hence, n order to have the hghest spectral effcency, each ase staton n each cell has to change the rato of uplnk and downlnk traffc volumes adaptvely for each data frame. Adjustng the uplnk and downlnk volume swtchng pont n each cell may rng hgher spectral effcency for the entre network, ut t also creates another prolem that should e resolved. By havng dfferent swtchng ponts for uplnk/downlnk n dfferent co-channel cells, D-TDD schemes are serously lmted y co-channel nterference (CCI) from ase statons n other co-channel cells. Ths s ecause the propagaton from a ase staton to a ase staton suffers less attenuaton than that from suscrer antenna to a ase staton ecause the antenna heght at a ase staton s much hgher than that of a suscrer antenna. Jeong et al. [27] ntroduced a tme slot allocaton algorthm mnmzng the outage performance wthout consderng rate, power, and andwdth allocaton. Whle the analyss s ased on perfect knowledge of other ase statons frame resource nformaton such as the swtchng pont and actve sector ndex, t mght e dffcult to otan perfect frame nformaton of all co-channel cells n a practcal stuaton ecause every cell wll allocate tme slots for etter spectrum effcency after they roadcast frame nformaton. Hence, we need to develop a tme slot allocaton algorthm that does not need perfect frame resource nformaton ut rather oundary nformaton.

73 60 In ths chapter, we present a statstcal assessment of co-channel nterference of D-TDD scheme when swtched eam sector antennas are used at ase statons, and we also present a cooperatve co-channel nterference avodance algorthm n TDMA/D- TDD fxed wreless cellular systems. The focus s on the nterference durng the uplnk cycle ecause uplnk tme slots experence more severe co-channel nterference from other ase staton (BS) antennas. In secton 5.2, we refly descre the system model and the co-channel nterferers, and we otan dstruton of co-channel nterference. In secton 5.3, we explan the co-operatve tme slot allocaton algorthm, and we show numercal results n secton 5.4. Numercal results show the co-operatve algorthm mproves the sgnal-tonterference rato (SIR) outage proalty. 5.2 System Confguratons Frame Format We consder a fxed wreless cellular system wth D-TDD/TDMA. Each cell s hexagonal, and all users are consdered to e located randomly wth a unform dstruton. As n Fgure 5-1, each frame has a contenton perod so that each suscrer sends a transmsson request to the BS t elongs to. The BS determnes how many tme slots wll e used for uplnk/downlnk and roadcasts the results. Based on the shared nformaton aout the oundary etween up/down lnks n every co-channel cell, each ndvdual cell wll try to allocate extra downlnk tme slots and extra uplnk tme slots,

74 61 avodng possle severe co-channel nterference. Fgure 5-1 shows an example of a D- TDD/TDMA frame havng a maxmum numer of extra tme slots equal to sx. For smulaton, we assume we have maxmum numer of extra tme slots to e 6 or 12. Fgure 5-1: D-TDD/TDMA frame structure As reported n [27], swtched eam sector antennas are deployed at the BS ste for nterference avodance. At each sector, a square aperture antenna s mounted such that the half power eam wdth (HPBW) ponts of antennas are overlapped. Hence, the numer of sector antennas equals to 360 /HPBW. In ths chapter, we use 8 (18) sector antennas wth an HPBW of 45 (20 ). Also, square aperture antennas are deployed at suscrer stes wth the same antenna pattern as n [44] 1+ cosφ κa E ( φ ) = c snc sn φ, where c s a constant, κ s the wave numer, a s the lateral sze of the aperture, and φ s the azmuth angle. The power gan pattern s gven y ( φ ) 20 log10 ( E( φ ) ) G.

75 5.2.2 Traffc Model 62 In the D-TDD system analyzed heren, we assume the frame accommodates all uplnk tme slot requests (n Up ). The frst and the last L-tme slots are assgned to uplnk receptons and downlnk transmssons. Durng ths tme slot nterval, all the co-channel cells are n the same cycle as n the reference cell. We denote ths tme slot nterval as the fxed tme slot regon, Π Fx. The rest of the tme slots n the frame are assgned ether to the uplnk or to the downlnk transmssons (otherwse, they reman dle) accordng to the traffc load. The tme slot nterval from the (L+1)-th to the last uplnk tme slot s called the extra uplnk tme slot regon, denoted as Π Ext. The spectral effcency of the TDMA system depends on the numer of actve tme slots: the spectral effcency would e mproved as the actve tme slots are ncreased. However, the ncrease of actve extra uplnk tme slots results n a strong BS-to-BS co-channel nterference, whch severely degrades SIR outage performance. Thus, we are only nterested n the effect of traffc varaton on the spectral effcency of D-TDD systems. The offered load for uplnk stream n a D-TDD frame λ UP s assumed to have mn max truncated Gaussan dstruton: λ ~ N( µ, σ n λ ). UP UP UP UP UP nup Hence, the numer of uplnk tme slots for D-TDD frame can e otaned as D nup λup =, 5.2 where selects the largest nteger whch s closest to the argument.

76 Assumng full-load condton, n D Up reflects the varaton of offered traffc for uplnk, and t also represents the swtchng pont from uplnk to downlnk tme slots. Here, the proalty dstruton for the numer of uplnk tme slots can e otaned as 63 P( n UP Pr{ nup = k) = mn Pr{ n n UP = k} UP k 1 µ Q σ UP = mn nup 1 µ Q σ UP n UP UP max UP } k µ UP Q σ UP max nup µ Q σ UP UP, 5.3 where 1 Q( x) e 2π = x t 2 / 2 dt Co-channel Interference As we mentoned refly n the ntroducton, the nature of co-channel nterference s dynamc over extra uplnk tme slots n D-TDD systems. Fgure 5-2 shows some examples of nterferers for reverse lnk and forward lnk. Conventonally, suscrers n co-channel cells are man nterferers for reverse lnk, and ase statons n co-channel cells are man nterferers for forward lnk as n cases (a) and (). Wth the use of D-TDD, other ase statons can e the source of co-channel nterference over extra uplnk tme slot regon and, also, suscrers n other co-channel cells can e co-channel nterferers over the extra downlnk tme slot regon, as n cases (c) and (d). As each co-channel cell has ts own rato of uplnk/downlnk n D-TDD systems, for a certan tme slot, co-channel cells n downlnk cycle can e strong cochannel nterferers. Consequently, the statstcs of co-channel nterference are dfferent,

77 tme slot y tme slot, and get worse n uplnk tme slots near the swtchng pont of uplnk and downlnk tme slots. 64 Fgure 5-2: Example of co-channel nterferers When swtched eam sector antennas are deployed at BS stes, cumulatve dstruton functon of co-channel nterference at each sector can e otaned as P[ I ( t) x] = P[ M M k k Gj j + k D( t) j= 1 k U ( t) j= 1 k k g δ x] 5.4 j j where t s the tme slot ndex, D (t) / U (t) as sets of co-channel cells n the downlnk/uplnk cycle at tme slot t, respectvely. k Gj s the gan etween reference sector of reference cell and j-th sector of k-th co-channel cell and g k j s the average gan etween a sector of reference cell and users uploadng to j-th sector of k-th co-channel cell. k j s the proalty that sector j of co-channel cell k s ON for downlnk, k δ j s the proalty that sector j of co-channel cell k s ON for uplnk, and M s the numer of sector antennas at BS stes.

78 Assumng we know perfect frame nformaton of all co-channel cells and consderng nterference from adjacent co-channel cells only, we need to fnd 12 gan 65 matrces of sze M M k forg and g k, k = 1,2,, 6, n order to otan nterference statstcs ased on the geometrcal model n Fgure 5-3. Fgure 5-3: Geometrcal model of co-channel nterference when swtched eam sector antennas are deployed at BS stes G k j P G ( φ1) G ( φ ) 10 n D BB ξ /10 t t r 2 =, 5.5 π 2π φ 1 = k k = 1,2,,6, M 2π 2π φ 2 = mn π + φ1 ( j ), π φ1 + ( j ), 5.7 M M g k j = E R, Θ P tgt ( φ1) Gr ( φ 2 )10 n SB L ξ /10, j = 1,2,, M, 5.8

79 where expectaton s taken wth respect to r andθ r φ 1 = φ1 sn ( sn( φ2 + θ )), L 5.9 φ = π ( φ φ φ θ, ) π π where L = D + r 2Dr cos( φ 2 +θ ), 0 r R and θ, M M. We can regard the dstruton of co-channel nterference for each sector and each tme slot as Gaussan n db, fndng mean and varance, µ (t) and σ (t) y Schwartz and Yeh s method [45], where s the sector ndex, = 1,2,, M and t s the uplnk tme slot EU ndex, t = 1,2,, N + N. N s the numer of fxed uplnk tme slots, and N EU s the numer of extra uplnk tme slots. We provde some cumulatve dstrutons of SIR n each tme slot for omn-drectonal antennas and sector antennas at the ase staton n the numercal results secton. In a real stuaton, t may not e possle to know the exact frame nformaton of all co-channel cells ecause each BS wll perform TSA, smultaneously. Hence, we need to get approxmate statstcs of nterference, assumng the worst stuaton that co-channel nterference s drectly comng from co-channel cell BS, and the correspondng sector antenna s always algned wth reference BS ( φ 0 ). Hence, we can get the M 6 nterference matrx Ξ for each cell. The (,j)-th element of Ξ, ξ(,j), represents the nterference gan from j-th co-channel cell when the reference cell uses -th sector antenna. Ths nterference gan matrx can e used n a co-operatve TSA algorthm. 2 =

80 Co-operatve TSA Algorthm for D-TDD Systems Durng roadcast perod, each ase staton shares uplnk/downlnk swtchng pont nformaton etween adjacent co-channel cells. We assume that all nterference from co-channel cells comes from exact BS locatons. To avod complexty, we only allocate extra uplnk and downlnk tme slots. For extra downlnk slots, slot allocaton s done from the left end. Ths algorthm tres to fnd the most nterference-avodng user for the slot and then moves to the next rght slot untl all adjacent cells are n downlnk. For extra uplnk slots, slot allocaton s done from the rght end. Ths algorthm tres to fnd an optmal user among the pool, whch can avod the many downlnk cells nterference and then move to the next left slot untl all adjacent cells are n uplnk. Fgure 5-4 shows a smple example wth two cells. When user 4 n cell 1 s n uplnk cycle, the sector antenna for downlnk of user 2 of cell 2 ecomes a strong cochannel nterferer. Smlarly, when user 3 of cell 1 s durng uplnk cycle, the sector antenna for downlnk of user 1 n cell 2 ecomes a strong co-channel nterferer. In cell 1, n order to avod havng strong co-channel nterference from cell 2, user 3 and 1, user 4 and 2 need to swtch ther uplnk tme slots respectvely. In cell 2, n order to avod sendng strong co-channel nterference to cell 1, user 3 and 1, user 4 and 2 need to swtch ther downlnk tme slots, respectvely.

81 68 Fgure 5-4: Co-operatve TSA algorthm - two cells example The co-operatve tme slot allocaton procedure s as follows, N ED : numer of extra downlnk tme slots N EU : numer of extra uplnk tme slots N: numer of fxed uplnk tme slots N Emax : maxmum numer of extra tme slots U ED : user set for extra downlnk tme slots U EU : user set for extra uplnk tme slots D(t): set of downlnk co-channels at t-th tme slot U(t): set of uplnk co-channels at t-th tme slot π(t): user ndex for t-th tme slot s(l): sector ndex for user l ξ (, j) : nterference from j-th co-channel cell when the reference cell uses -th sector antenna

82 egn // allocate extra downlnk tme slots t N+ N Emax - N ED ; t max = N+ N Emax ; whle t < t max & U(t) φ π (t) mn{ ξ ( s( u, d) )} ; u U ED d U ( t) 69 U ED U ED {π(t)}; t t + 1; end // allocate extra uplnk tme slots t N+ N EU ; t mn = N+1; whle t > t mn & D(t) φ π (t) mn{ ξ ( s( u, d)) }; u U EU d D( t) U EU U EU {π (t)}; end end t t - 1; 5.4 Numercal results We assume a hexagonal cellular system wth a frequency reuse of 7 and a cell radus of 2km. Each TDD frame s dvded nto 48 tme slots, and 12 and 30(or 24) tme slots are dedcated to uplnk and downlnk, respectvely. Sx (or 12) extra tme slots are used for uplnk or downlnk ased on the asymmetrc traffc volumes for uplnk and downlnk. Thus, the maxmum rato for downlnk s 3/4 for oth cases. We have eght

83 70 sectored antennas wth an HPBW of 45. For smplcty, the channel s assumed to e flat fadng, and mult-path fadng s not consdered. Only SIR, ased on local mean power, s nvestgated as a performance measure for the systems, and no specfc modulaton scheme, error control codng, dversty or an equalzaton technque s consdered. For computer smulatons, we assume all users have packets to transmt at all tmes so that we could see the effect of strong co-channel nterference and TSA algorthms. Fgure 5-5 shows the CDF of SIR n each tme slot when omn-drectonal antennas are deployed at the ase staton. Tme slots n fxed regon (T 1 ~T 12 ) have the same statstcs of SIR; however, tme slots n the extra regon (T 13 ~T 18 ) show dfferent statstcs of SIR. Tme slots near to the swtchng pont of uplnk/downlnk have worse SIR statstcs. Fgure 5-5: CDF of SIR when omn-drectonal antennas are deployed at BS

84 71 Fgure 5-6 (a) and () shows CDF s of SIR n each tme slot when sector antennas are deployed at BS, efore and after tme slot allocatons, respectvely. From Fgure 5-6 (a), t s noted that the statstcs of SIR mprove wth sector antennas at ase statons even wthout tme slot allocaton. Those mproved SIR statstcs mprove more when we use tme slot allocaton for the purpose of co-channel nterference suppresson. Fgure 5-6: CDF of SIR when sector antennas are deployed at BS (efore and after TSA) Fgure 5-7 shows the SIR outage performance when the maxmum numer of extra tme slots s 6. The results are compared wth the SIR outage performance when omn-drectonal antennas are deployed at BS stes. As reported n [27], havng omndrectonal antennas at BS stes s ntolerale to the severe co-channel nterference from other BS antennas, whch s caused y dfferent swtchng ponts for uplnk/downlnk. Smulaton shows that the achevale SIR value wth 18 sector antennas at BS stes at an outage of 1% s aout 6 db hgher than that wth 8 sector antennas at BS stes. The computer smulaton also shows that TSA mproves achevale SIR values y aout 5 db at an outage of 1%.

85 72 Fgure 5-7: Aggregate SIR wth the 6 extra tme slots for 8 (18) sector antennas at BS Fgure 5-8 shows the SIR outage performance when the maxmum numer of extra tme slots s 12. Smulaton results show that the achevale SIR value wth 18 sector antennas at BS stes at an outage of 1% s aout 7 db hgher than that wth 8 sector antennas at BS stes. The computer smulaton also shows that TSA mproves y aout 8 db, the achevale SIR values at an outage of 1%.

86 73 Fgure 5-8: Aggregate SIR wth the 12 extra tme slots for 8 (18) sector antennas at BS Comparng Fgure 5-7 and Fgure 5-8, we fnd that the more we have extra tme slots, the severer co-channel nterference we experence from other co-channel ase statons. Ths s ecause of the fact that we need to reduce the numer of dedcated tme slots, whch have etter SIR statstcs than extra tme slots as showed n Fgure 5-6, to have more extra tme slots. 5.5 Conclusons In ths chapter, we evaluated co-channel nterference n a D-TDD system wth swtched eam sector antennas at BS stes and proposed a co-operatve TSA algorthm

87 74 that works n each co-channel cell ndependently, coverng an entre servce area. Each co-channel cell uses TSA algorthm smultaneously and allocates extra downlnk slots as well as extra uplnk slots, whch enhances co-channel nterference avodance. We also have seen that havng more sector antennas mproves avodng co-channel nterference and consequently, offers a etter outage performance. In the smulaton results wth 6 and 12 extra tme slots, we have seen that havng more extra tme slots results n more co-channel nterference. Therefore, t s of nterest to fnd the optmal numer of extra tme slots for the optmal throughput performance when the dynamcs of traffc s huge.

88 75 Chapter 6 ADAPTIVE RESOURCE ALLOCATION FOR TDD/TDMA SYSTEMS WITH HETEREGENEOUS TRAFFIC 6.1 Introducton It has een assumed that swtchng ponts for uplnk and downlnk of TDD schemes are decded y network operators and are not changeale. Moreover, swtchng ponts n adjacent cells have een synchronzed n order to avod severe nter-cell nterference. Havng the same swtchng ponts n adjacent cells, a centralzed controller could set swtchng ponts for all cells y oservng the traffc characterstcs [43]. In ths chapter, we propose a cross-layer desgn of dynamc TDD (D-TDD) scheme consderng the traffc and channel condton together. Snce each cell has dfferent offered loads for uplnk and downlnk, we set swtchng ponts of all cells, ndependently. Ths may cause severe co-channel nterference around the swtchng ponts, ut t stll produces more throughput than the conventonal fxed TDD (F-TDD) scheme. Swtchng ponts can e updated on a daly ass or a frame ass. After montorng for a certan amount of tme, a network operator may update the swtchng pont, and resources are allocated ased on the swtchng pont nformaton. In ths

89 76 chapter, we update the swtchng pont on a frame ass, changng t dynamcally for each frame ased on the traffc characterstcs and channel status. Based on the resource allocaton algorthm, each user wll e gven a numer of tme slots wth ndces and a modulaton format. The modulaton s decded y the estmated channel state and the ndces of tme slots are decde y the tme slot allocaton (TSA) algorthm. Chung et al. [16] reported that y usng just one or two degrees of freedom n adaptve modulaton yelds somethng close to the maxmum possle spectral effcency otaned y utlzng all degrees of freedom. Moreover, n cellular archtecture, power adaptaton could nduce hgher nterference at co-channel cells. Hence, we do not deal wth power adaptaton n ths chapter. The remander of ths chapter s organzed as follows. We descre our system n the followng secton, and ts performance analyss wll e descred n secton 6.3. Secton 6.4 addresses some performance evaluaton crtera. Secton 6.5 presents some numercal results and, fnally, the chapter s summarzed and concluded n secton System Confguratons Ths secton provdes the confguraton of D-TDD/TDMA systems for adaptve resource allocaton.

90 6.2.1 Frame Format 77 Fgure 6-1 shows the frame format of our D-TDD systems, whch s slghtly dfferent from the frame format n Chapter 5. Ths s ecause we have a dfferent traffc model n ths chapter; we have two classes of servce wth dfferent prorty levels and andwdths. New connectons and hand-off connectons are requested through contenton slots. Feedack from on-gong calls wll e pggyacked n uplnk packet va an uplnk channel. Fgure 6-1: Frame format of D-TDD system wth two classes of servces Each ase staton (BS) decdes the uplnk/downlnk oundary ased on the requested aggregate data rate for uplnk and downlnk. Through the roadcast perod, ths oundary nformaton s roadcasted to BS s n co-channel cells and then, the result of tme slot allocaton (TSA) s roadcasted to each remote termnal wthn each cell.

91 6.2.2 Channel Model 78 The average receved sgnal power level, P r, s a functon of the dstance etween the transmtter and recever, R, the path loss exponent η, the transmtter power, P t, and the transmtter and recever antenna gan, G andg, respectvely, as t r PG t tgr Pr. 6.1 η R Havng fxed values for the transmtter power, the transmtter and recever antenna gan, the average sgnal-to-nose rato, γ, can e represented as a functon of the dstance etween the transmtter and the recever 1 γ = κ, 6.2 η R where κ s a constant dependng on the gan and heght of the antennas and the carrer frequency. The path loss exponent η n ths chapter s consdered to have values of 4 and 3 for propagaton etween ase statons and etween ase staton and suscrers, respectvely. Now, the nstantaneous recever SNR can e expressed as where ξ s a Gaussan random varale for long-term attenuaton, shadowng, and α s a Raylegh random varale or a Rce random varale for short-term attenuaton. ξ 10 2 α γ = κ 10, 6.3 η R

92 6.2.3 Adaptve Modulaton 79 Fgure 6-2 shows an example of roadand wreless access (BWA) systems wth adaptve modulatons. Fgure 6-2: An example of adaptve modulaton for BWA We do not deploy power adaptaton ecause t can nduce hgher nterference at co-channel cells n a cellular archtecture wth frequency reuse. However, we set the (T ) transmtter power from the ase staton so that the target packet error rate (PER), P, can e satsfed at the edge of cells. We assume multple modulaton formats can e used ased on the estmated channel state nformaton, whch can e provded y the recever va a feedack channel. We have two dfferent classes of calls as n Tale 6-1 n secton We change modulaton format for calls for class B only, always satsfyng the target PER.