Resource Allocation in Multi-cell OFDMA-based Relay Networks

Transcription

1 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. Resource Allocaton n ult-cell OFDA-ased Relay Networks Yao Hua Tsnghua Natonal Laoratory for Informaton Scence and Technology, Department of Electronc Engneerng, Tsnghua Unversty, Beng Chna huay@mals.tsnghua.edu.cn Qan Zhang Department of Computer Scence, Hong Kong Unversty of Scence and Technology qanzh@cse.ust.hk Zhsheng Nu Tsnghua Natonal Laoratory for Informaton Scence and Technology, Department of Electronc Engneerng, Tsnghua Unversty, Beng Chna nuzhs@tsnghua.edu.cn Astract Cooperatve relay networks comned wth Orthogonal Frequency Dvson ultplexng Access (OFDA) technology has een wdely recognzed as a promsng canddate for future cellular nfrastructure due to the performance enhancement y flexle resource allocaton schemes. The maorty of the exstng schemes am to optmze sngle cell performance gan. However, the hgher frequency reuse factor and smaller cell sze requrement lead to severe nter-cell nterference prolem. Therefore, the mult-cell resource allocaton of sucarrer, tme schedulng and power should e ontly consdered to allevate the severe nter-cell nterference prolem. In ths paper, the ont resource allocaton prolem s formulated. Consderng the hgh complexty of the optmal soluton, a two-stage resource allocaton scheme s proposed. In the frst stage, all of the users n each cell are selected sequentally and the ont sucarrer allocaton and schedulng s conducted for the selected users wthout consderng the nterference. In the second stage, the optmal power control s performed y geometrc programmng method. Smulaton results show that the proposed the nterference-aware resource allocaton scheme mproves the system capacty compared wth exstng schemes. Especally, the edge users acheve more eneft. I. INTRODUCTION Due to the attractve features of hgh throughput, low power consumpton and extended coverage, the cooperatve relay network archtecture has the trend to replace the tradtonal pont-to-mult-pont network archtecture. eanwhle, OFDA technology s ecomng a popular choce snce the frequency selectve fadng prolem n roadand system can e allevated y dvdng the whole andwdth nto multple parallel sucarrers. The ncorporaton of OFDA technology and the relay network structure provdes a promsng platform, whch offers nce flexlty n terms of resource allocaton, such as sucarrer allocaton, schedulng and power control to acheve the mult-dmensonal dversty gan. Therefore the emergng G wreless system such as Long Term Evoluton- Advanced (LTE-A) [] and Wax [] wll adopt such a promsng OFDA-ased relay network nfrastructure. The resource management s crucal to guarantee the system performance of a wreless system. In OFDA-ased relay networks, most of exstng works concentrated on the sngle cell scenaro and share the asc dea as allocatng resources to maxmze the local performance gan [][][][3]. However, n future wreless networks, smaller cell sze and hgher Fg. : Jont resource allocaton scheme frequency reuse factor s mperatve to meet the growng capacty requrement, whch nduces severe nter-cell nterference prolem. The nteracton among adacent cells results n the fact that the maxmzaton of local performance gan may not acheve the gloal optmal performance. Take the scenaro n Fg. as a motvatng example. Two relay lnks n two adacent cells are consdered. The lnks etween ase statons (BSs) and relay statons (RSs) are defned as ackone lnks and denoted y sold arrows. The lnks etween RSs and mole statons (Ss) are defned as access lnks and denoted y hollow arrows. Two avalale sucarrers are represented y two colors, where the red channel has hgher channel gan n all of the lnks. We name Scheme as the one that tres to maxmze the local performance gan,.e., oth of the two relay lnks select the red channel. However, the lnks suffer strong mutual nterference and proaly Scheme cannot acheve optmal sum throughput. Snce another lue channel s avalale, although the channel gan s lower than that of the red channel, allocatng orthogonal channels to neghorng lnks can effectvely reduce the nterference, such as Scheme. The analyss aove reveals that the key ssue of optmal resource allocaton s to acheve the tradeoff etween local performance gan and the mpact of nterference. To ontly consder the two factors, the sucarrer allocaton, schedulng and power control should e ontly consdered //$6. IEEE

2 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. Frst, sucarrer allocaton and schedulng should e ontly consdered. When the same sucarrer s allocated to the lnks n neghorng cells, f the schedulng s not ontly consdered, nterference from neghorng BSs may degrades the throughput more severely snce BSs have hgher transmsson power than RSs. On the other hand, due to the two-hop nature of relay lnks, approprate schedulng order may avod the domnant nterference from BSs. For the example shown n Fg., we can have Scheme 3, whch swaps the schedulng orders of access lnks and ackone lnks of the two cells so that the domnant nterference from BS (e.g., B ) to the neghorng RS (R ) s susttuted y the much weaker nterference to the neghorng S (U ). What s more, the power control scheme should e also coupled wth sucarrer allocaton and schedulng, ecause oth the sucarrer and tme slot allocaton determne whch lnks nterfere wth each other and how much power should e allocated to allevate the nterference. Once the sucarrer allocaton and schedulng s fxed, the power allocaton of BSs and RSs can e used to allevate the nterference, such as Scheme of Fg.. Reversely, f lnks are too close to each other and the nterference can hardly e allevated y power control, orthogonal resources n ether frequency (Scheme ) or tme doman (Scheme 3) should e allocated to avod the nterference. In ths case, the nterference s avoded at the cost of local performance degradaton. Therefore, the ont resource allocaton scheme s mperatve to acheve the optmal tradeoff etween the two factors. The ont resource allocaton prolem can e formulated as a mxed nteger programmng (IP) prolem [7], whch s NP hard. The man contruton of the paper s that a lowcomplexty ont sucarrer allocaton, schedulng and power control scheme s proposed. The scheme s dvded nto two stages to reduce the complexty. In the frst stage, one S s selected n each cell n each round ased on the channel state nformaton (CSI). Then the sucarrer allocaton and schedulng scheme s operated for the selected S. Ths operaton terate untl all of the sucarrers have een allocated. In the second stage, the power control s conducted to optmze the throughput for all of the allocated sucarrers. The very nce oservaton we have s that when the Amplfyand-Forward (AF) relay technology s adopted, the mult-cell power control prolem can e optmally solved y geometrc programmng method n the hgh Sgnal to Interference Rato (SIR) regon. Smulaton results demonstrate that the proposed resource allocaton scheme sgnfcantly enhances system capacty, especally for the edge users. The rest of ths paper s organzed as follows. In Secton II, the related work s summarzed. Then a ont resource allocaton prolem s formulated n Secton III. After that an teratve resource allocaton scheme s proposed n Secton IV. Secton V shows the smulaton results. Fnally, Secton VI concludes the paper. II. RELATED WORK Numerous prevous work nvestgated resource allocaton schemes n OFDA-ased relay networks [][][][][3][7]. The sucarrer allocaton and the schedulng scheme were adopted to acheve frequency and tme dversty gan. However, most of the prevous work focused on sngle cell scenaro and overlook the nterference among adacent cells, whch s the system capacty ottleneck n future cellular networks. What s more, the power control scheme was not addressed, whch plays an mportant role for nter-cell nterference allevaton. Some power control schemes have een exploted for nterference mtgaton n relay networks [6][]. These work consdered that several relay lnks transmt smultaneously and all of the lnks use the same frequency. However, as we have mentoned n Secton I, power control should not e separated from the sucarrer allocaton and schedulng. In [], the pont-to-mult-pont scenaro was dscussed and a mult-user power control prolem was solved. All of the lnks n the cell used orthogonal resources and the nter-cell nterference prolem was not addressed. There have also een some work on the ont sucarrer and power allocatons schemes of OFDA-ased relay networks under mult-cell scenaro[7][8]. [7] proposed a dstruted sucarrer and power allocaton scheme. Each cell ndvdually decdes the resource allocaton wthn ts cell accordng to the measurement of nterference from neghorng cells. However, the cooperatve data s relayed y the adacent BSs. In that case, the relay transmssons occupy the spectrum resource of the relayng BSs, therefore the system spectrum effcency s lower than that wth specfc RSs. What s more, due to the dstruted property of the resource allocaton scheme, the nterference was assumed constant, whch s not practcal when the transmsson powers of the nterferng cells are adusted. In [8], a ont resource allocaton prolem n mult-cell scenaro was also conducted. However, the per-tune power constrant n each cell was assumed, whch s not practcal. In [], the geometrc programmng (GP) method s utlzed to solve the power control prolem n tradtonal cellular networks. otvated y ths work, we proved n ths paper that when the AF relay technology s adopted, n the hgh SIR regon the smplfed power allocaton prolem can e optmally solved y the GP method and the optmalty of the proposed resource allocaton scheme can e guaranteed. To summarze, the severe nter-cell nterference prolem rngs new challenges to resource allocaton of future OFDA-ased relay networks. The ont allocaton of sucarrer allocaton, schedulng and power control should e proposed to acheve the gloal optmal performance. III. JOINT RESOURCE ALLOCATION PROBLE A. System odel Consder a mult-cell wreless networks wth several dedcated relay statons n each cell. There are totally N BSs. For each BS B n (n [,N]), there are R RSs and Ss elongng to the cell. Snce BSs can cooperate wth each other more easly y sgnallng transmtted va wred core networks, we focus on downlnk relay transmsson. The Ss wth poor drect lnk channel state are frst selected for

3 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. relay transmsson y some rough crterons, whch wll e addressed n the next susecton. Then the resource can e dvded nto two portons for drect and relay transmssons respectvely. We only focus on resource allocaton for the selected Ss whch have determned ther transmsson modes (drect mode or relay mode). The fxed AF relayng strategy wthout recever comnaton s used [8]. Each fxed AF cooperatve transmsson s dvded nto two slots. The duraton of the two slots are dentcal. In the frst slot, only BS transmts and RS receves. In the second slot, only RS transmts and S receves. We assume that the transmsson among dfferent BSs are also synchronzed,.e., the ntra-frame nterference s not consdered. Such synchronzaton can e acheved y adoptng Gloal Poston System (GPS) as used n TDD-LTE system []. All of the Ss are assumed fxed and the channel gan s constant durng each frame. For practcal mplementaton, each S s allowed to communcate wth at most one RS to reduce the complexty of synchronzaton wth multple RSs at the physcal layer. We assume that the relay selecton s done efore the resource allocaton scheme accordng to long term CSI, such as [6]. Because wthout such assumpton the onlne resource allocaton for the relay nodes cannot e realzed. Also we regulate that the sgnals from BS and RS cannot e comned due to synchronzaton dffculty. Snce we have fxed the relay selecton for each S, smart resource allocaton scheme can e adopted to mnmze the nterference among the system. There are three resources that can e allocated for nterference mtgaton. The frst s the transmsson power. We assume that each sngle BS has the maxmum overall transmsson power P max constrant of all the sucarrers allocated to the BS. Each sngle RS has the maxmum overall transmsson power P rmax constrant of all the sucarrers allocated to the RS. The transmsson power of B n to the th RS R and the transmsson power of R on sucarrer k s denoted y P k to the th S n the cell on sucarrer k s denoted y P. The second resource s the sucarrer. We can dynamcally decde whch of the sucarrers are allocated to whch access/ackone lnks accordng to the nstaneous CSI. We assume there are totally K sucarrers avalale n each cell. Each sucarrer s allowed to e used only once n all of the access lnks and only once n all of the ackone lnks of one cell respectvely. Although some references showed the potental performance gan y allocatng the same sucarrer to dfferent access lnks [3], we proht ths allocaton n our scheme. Further we assume that the data transmtted va one sucarrer n the ackone lnk to any RS can e only forwarded to one S wth one sucarrer,.e., each sucarrer cannot e shared y more than one data stream. We denote as the ndex of allocatng the sucarrer k to the access lnk from R to and r as the ndex of allocatng the sucarrer k to the ackone lnk from B n to R. The last resource s the schedulng order of the two-hop relay lnks. Snce some data should e transmtted va two hops a to the Ss usng the relay mode, each stream elongng to any partcular S can decde whether the access lnk s scheduled n the frst slot or the ackone lnk n the frst slot. As we have demonstrated n Secton I, the approprate schedulng order can effectvely avod the nter-cell nterference. We use to ndcate the schedulng ndex of the access lnk to the th S n the cell n. y ndcates the schedulng ndex of x the ackone lnk to the th RS n B n. All of the schedulng ndexes are {, } value, where represents schedulng the correspondng lnk n the frst slot and represents n the second slot. B. Prolem descrpton and Formulaton The ont sucarrer allocaton, schedulng and power control prolem s formulated n ths susecton. Snce the prolem s comnatoral, a low-complexty two-stage soluton s proposed n the next secton. We frst concentrate on the S n the cell B n and ts assocated RS R. To evaluate the nterference the two-hop lnk suffers, we frst assume that the sucarrer k and k has een allocated to the access lnk and ackone lnk of respectvely. The channel gan from any BS B n to R and at channel k are denoted as G (n n) sr () and G (n n) sd (). The channel gan from RS R (n ) to R and at channel k s denoted y G (n n) rr ( k ) and G (n n) rd ( k ) respectvely. Then the receved SIR of the drect lnk α, ackone lnk and access lnk γ can e expressed y Eq.(), (), and (3) respectvely. B s the andwdth of each sucarrer and N s the whte nose power. The suscrpt k represents that the access lnk adopts the sucarrer k and the matched sucarrer k n the ackone lnk s omtted. Before the ont resource allocaton scheme s operated, each S should frst decde whether the drect mode or the relay mode should e selected for transmsson. Snce the transmsson mode decson should e smple to reduce the complexty of the scheme, n ths paper, we smply use the nstaneous SIR for the mode decson as follows: D, max k K R, else α > max k,k K γ + γ + where D s denoted as the set of Ss usng the drect lnk and R as the set of Ss usng the relay lnk n the nth cell, respectvely. Snce the fxed AF relay technology wthout recever comnaton s adopted, the recevng SIR of at sucarrer k s: SIR = γ + γ +, R α, D, () ()

4 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. α = rk n N P (n ) k G (n n) sd ( ) y (n ) [ R(n ) n =,n =n = = rk n N P (n ) k G (n n) sr ( ) y (n ) γ [ R(n ) n =,n =n = n =,n =n = P k G (nn) x sd ( ) + (n ) = P k G (nn) y sr ( ) + (n ) = = a n P G(nn) rd () N [ R(n ) P (n ) n) k G(n sd ( ) y (n ) x + (n ) =, () P (n ) k G (n n) rd ( k ) x (n ) x ]+N, () P (n ) k G (n n) rr ( k ) x (n ) y ]+N. (3) P (n ) n) k G(n rd ( k) x (n ) x ]+N and the data rate acheved y R = K k= s: log ( + SIR ). (6) Then, our optmzaton goal of the ont resource allocaton prolem s to maxmze the overall throughput of the system: maxmze N n= = R (7) The frst constrant of the optmzaton prolem s that each sucarrer k can e used y only one of the ackone lnks or one of the access lnks n each cell,.e.: C: = R = a, n,, k r k, n,, k a,r k {, }, n,, k Also, there are maxmum transmsson power constrants for each BS and RS,.e.: C: R K = k= = K k= P k P max, n P P rmax, n, Last, the schedulng ndexes should e {, } varale, and each par of sucarrers allocated for one relay lnk cannot e scheduled smultaneously,.e.: C3: x + r =, n,, : x,y {, } n,, R (8) (9) () IV. IARA: AN DISTRIBUTED SUBCARRIER ALLOCATION, SCHEDULING AND POWER CONTROL SCHEE In Secton III, the optmal resource allocaton n multcell relay-ased OFD networks s formulated. However, the soluton of the optmzaton prolem of Eq.(7) wth constrants Eq.(8), (9) and () s NP hard. It has een proved y Theorem 3 n [] that, the ont sucarrer allocaton and schedulng prolem n the sngle cell relay network s NP hard. The computaton complexty of our prolem s at least N tmes that of the sngle cell prolem even wthout the power control. Therefore, the total computaton complexty of the optmzaton prolem s NP hard as well. Snce the computaton complexty s extremely hgh, a lowcomplexty suoptmal resource allocaton scheme s proposed n ths secton. The resource allocaton scheme s dvded nto two stages to reduce the complexty. In the frst stage, a sucarrer and schedulng scheme s proposed. In the frst stage, one par of sucarrers s allocated n each cell and ther schedulng order s decded. Ths process terates untl all the sucarrers have een allocated. By such operaton, all of the nteger varales are fxed. In the second stage, a power control scheme s constructed to optmze the contnuous power varales for all of the allocated sucarrers. An geometrc programmng (GP) method s used to derve the optmal power allocaton result. To further reduce the computaton complexty, the dual decomposton method can e adopted n the power allocaton stage such that the prmal prolem can e decomposed nto N suprolems, each of whch s solved y one BS and only requres local CSI. The computaton complexty of sucarrer and schedulng strategy s delectale compared wth power allocaton. The computaton complexty of orgnal power allocaton prolem s O(K(N + RN)). By decouplng the prolem nto separate dual prolems, the complexty s O(K), whch s reduced y one order of magntude compared wth the prmal prolem. A. Sucarrer Allocaton and Schedulng Strategy The ont scheme starts wth the teratve sucarrer allocaton and schedulng decson. In each round, ased on

5 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. the current nterference level, each cell greedly selects one partcular S and allocates one par of sucarrers and the schedulng order to maxmze the throughput of the selected S, where the maxmum transmsson powers of the BS and RSs of the selected S are assumed. For each par of sucarrer allocaton, the two dfferent schedulng orders are oth tred. Fnally, the greedy sucarrer allocaton and schedulng decson s made as follows: (k, k ) = argmax R,,x,y n, k C a,k C () where C a and C are the set of sucarrers that have not een allocated for the access lnks and ackone lnks n cell B n respectvely. After allocatng sucarrer k and k, ths par of sucarrers can e removed from C a and C respectvely. By operatng the same procedure for all of the N cells, N Ss wth N sucarrer pars wth ther schedulng order have een allocated. After K rounds of allocaton, all the C a and C ecome empty and the teraton termnates. Notce that all the N sucarrer allocatons are selected n parallel wthout consderng ther mutual nterference. After decdng the allocated sucarrers and ther schedulng order, the optmal power control s operated as the next susecton descres. B. Power Control: A Geometrc Programmng ethod As we have mentoned n Secton I, the mult-cell power allocaton scheme can e solved wth the geometrc optmzaton method after all of the dscrete varales n Prolem (7) are fxed. To further reduce the computaton complexty, the dual decomposton method s also used to dvde the orgnal prolem nto N suprolems. After the frst stage of sucarrer allocaton and schedulng, the set of sucarrers allocated to the access and ackone lnks, namely S a and S, are fxed for each cell B n and ther correspondng schedulng ndexes x and y are also determned. Then the power control prolem s transformed as Eq.() shows. In ths optmzaton prolem, only transmsson powers are varales. However, due to the exstence of nterference tems n the denomnator, the optmzaton prolem s non-convex, whch makes the soluton extremely dffcult. However, we oserve that once all of the Ss have determned ther transmsson modes, n hgh SNR scenaro the prolem descred y Eq.(7) can e approxmately solved y GP method under the followng smplfcatons: snce each S always chooses the lnk wth hgher channel condton for transmsson, the SIR of the selected sucarrer s often much larger than one and the rate expresson of R can e reformulated as Eq.(3) demonstrates. I and I denote the set of lnks n the system whch have een allocated the sucarrer k and transmt smultaneously wth the lnk G r () and G rd (). P m s the transmsson power allocated to the correspondng sucarrer m. maxmze s.t. N n= k S a = = k S a R P k a P rmax, n, R = k S P,P k P r k P max, n, n,,, k. () It s oserved from Eq.(3) that, no matter whch transmsson mode s used, f the recevng SIR s large enough, the n the log of Eq.(3) can e omtted. Also, the n the denomnator of SIR expresson of the relay lnk can e omtted. Then the reverse of the expresson n the log of Eq.(3) s a posynomal respect to transmsson powers. Therefore, the optmzaton goal of Eq.() can e reformulated to the mnmzaton of products of a seres of posynomals, whch s stll a posynomal as Eq.() shows. Consequently, the power control suprolem can e solved y GP method. mnmze N K ( n= k= D α R +γ γ ) () Fnally, snce the centralzed power control prolem s tme consumng, the dual decomposton method can e utlzed for the GP method. As descred y Secton V of [], the auxlary power P s ntroduced, where ndex represents the nterfered BS and represents the nterferng BS, k represents the partcular sucarrer. The dual of the orgnal prolem s as Eq.(), where S = S a S. Ovously, the dual prolem can e dvded nto suprolems for each sucarrer and the computaton complexty can e sgnfcantly reduced. The suprolem of BS can e formulated as Eq.(6). Notce that The soluton of each dual prolem only requres the local CSI,.e., the CSI of nterferng lnks to tself and the ntra-cell CSI, whch can e receved y neghorng cell sgnallngs. () mn L =sup P N =, = = k S () a k S () λ P + m I () P m G m + N G () k P () k N λ P () k k S () = (6) After each teraton, the dual costs of the axllary power tems are updated to solve the prmal prolem. The process terates untl the power allocaton result converge to the stale pont. Snce the oectve functon s dfferentale and feasle solutons exst, the sugradent proecton method can e utlzed [9] to otan the optmal prmal prolem soluton. The dual cost can e updated as follows:

6 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. R log α = log R log γ + γ G (nn) s ()P m I P m G m + N, G (nn) r = log N ( m I ()P D, k G(nn) rs ()P P m G m )G (nn) r ()P + N ( P s I sk G sk)g (nn) rs, R, k ()P (3) mn L =sup P N n= = R k S N N S () = =, = k= λ [ P P () k ] () The Procedure of IARA : Intalze: S a and S s set the total numer of sucarrers for each cell : Each of S ndvdually decde whether to use drect lnk or relay lnk ased on Eq.() 3: for n =;n N; n = n + do : Select one par of sucarrers (k,k ) for one S n cell B n followng Eq.() : S a =S a / k, S =S / k 6: end for 7: f S a, n [,N] then 8: Go to 3 9: end f : Do power control for all the allocated sucarrers as Secton IV-B descres : Exts λ (n +) = λ +γ[ P P ],,, k S () (7) where γ s the a step sze sequence le γ = n to grantee the convergence of the sugradent proecton method. Snce the transformed GP prolem s convex, the converged power allocaton result s also optmal. C. Jont Power,Sucarrer and Schedulng Allocaton Scheme In ths susecton, an Interference-Aware Resource Allocaton (IARA) scheme s performed for mult-cell OFDAased relay networks. In the frst stage, an teratve sucarrer allocaton and schedulng s operated, as Secton IV-A shows. After K rounds of teraton, all of the sucarrers are allocated to each cell and the teraton termnates. Then, power control s operated on all of the allocated sucarrers as Secton IV-B descres. The process can e shown y the followng tale. D. Implementaton Issues In ths secton some key mplementaton ssues of IARA are dscussed. Snce IARA should e operated for each OFD frame, the transmsson process durng each frame perod s descred here and the mplementaton ssues wll e ntroduced as the order of the transmsson process. Before the cooperatve transmsson, the followng two operatons should e done: the CSI measurement and the relay selecton. For the CSI measurement, BSs acqure the CSI of oth access lnks and ackone lnks among the system for ont resource allocaton. BS frst roadcasts the plot packets for channel estmaton. After recevng the packets, the RSs/Ss perform channel estmaton usng the OFD plot symols n the packet header. For each frame (also the resource allocaton perod) eng consdered, each of the Ss wthn the cell returns the measurement result y dedcate uplnk channel and fnally Ss are scheduled for cooperatve transmsson n ths OFD frame. The S schedulng scheme can refer to any lterature aout mult-user schedulng of OFD system, such as [3]. The tme granularty of the relay selecton scheme should e much larger compared wth the resource allocaton scheme ecause the large tme scale resource allocaton should e done ased on the selecton result efore we explot the opportunstc gan of reducng nterference n small tme scale. Relay selecton method adopted n ths paper refers to [7]. Each S compares the average channel gan to all of the RSs and select the hghest one as ts RS. Also each S can decde whether to use drect lnk or relay lnk accordng to the local CSI. After the CSI acqurement and the relay selecton, the proposed resource allocaton scheme IARA can e carred out. The sucarrer and schedulng allocaton stage does not requre any nter-cell CSIs. In the power allocaton stage, the dual decomposton s utlzed such that only local CSI nformaton s requred, for power computaton, whch guarantees the scalalty of the scheme. For each BS, the CSI of the cells wth nterference powers aove some threshold P th are

7 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. recorded and made nput for the power allocaton. All of the dual costs are gven any postve ntal values. Then the decomposed prolem of Eq.() s solved y each BS. After that the calculated optmal power s dssemnated to all of the BSs whch have strong enough nterference powers for the dual cost update y Eq.(7). After some teratons of dual cost update, the optmal power allocaton can e acheved. After all the of the resource allocaton have een done, the OFD frame can e transmtted. In cooperatve transmsson phase the recever may experences a tme offset etween the sgnals receved from the ts own cell and out of the cell. Therefore the key mplementaton dffculty s the the mult-cell frame level synchronzaton. The OFD frame synchronzaton can e solved y the scheme n [9], where the cyclc prefx (CP) of the OFD frames can e used to synchronze the tme offset of dfferent frames. We should regulate all of the BSs n the system to adopt the same synchronzaton pattern such that the S can synchronze the frames from other cells. V. PERFORANCE EVALUATION In ths secton we adopt dscrete tme event smulator to verfy the convergence of the teratve power control scheme and the effectveness of the proposed IARA ont resource allocaton scheme. A. System Setup Our smulaton s ased on the followng topology: twoter cells are consdered where the cell radus of each cell s 7m. Three RSs are evenly located n each cell, as Fg. shows. Equal numer of Ss are randomly dstruted n each cell. Each selected S has saturated traffc demand. The Okumura-Hata path loss model s adopted []: l(d) = log(d) n db (where the unt of d s Km). The smulaton parameters are lsted n Tale.I. The..d lock Raylegh fadng channel model s utlzed. The channel condtons keep constant durng each OFD frame and vary n dfferent frames. The frame duraton T s mllseconds. The default value of path loss α s and shadow standard devaton δ s 7 db, whch are the typcal values for outdoor fadng envronment [7]. We should frst valdate the convergence of the proposed dstruted scheme to guarantee the scalalty. Specfcally, the convergence speed of the dual GP power control s analyzed. Then the accuracy of the proposed power control scheme s valdated. After that, the throughput performance varaton wth dfferent network denstes s gven to oserve the performance gan of IARA n dfferent wreless envronments. The most related work n [7] s compared. Snce the orgnal work of [7] adopts the mult-bs cooperaton, for the purpose of comparson, the scheme n [7] s adapted nto the topologes wth specfc RSs. The performance of IARA s also compared wth the optmal result, whch s otaned y exhaustve search. Fnally, the performance gan of the edge Ss are dscussed specfcally to show the advantage of the relay archtecture. 6 7 R r Fg. : Smulaton topology. TABLE I: Smulaton Parameters parameters value unts N 9 R 3 K 8 R 7 m r 3 m B Hz α δ 7 db P max W P rmax 3 W N - dbmw T ms Notce that although the dual decomposton s used to reduce the computaton complexty y one order, the complexty s stll too hgh to e acceptale for real tme algorthm. The rest of the results shows that y ntroducng the nterferenceaware resource allocaton schemes, the OFDA-ased relay networks can endure more nterference and acheve hgher cooperatve gan. B. Smulaton Results We frst verfy the convergence of the proposed power control algorthm. A -BS topology (cell and n Fg.) s adopted. For smplcty we concentrate on the allocated powers of BS and the three RSs n cell. We run the dual decomposton method of Eq.() for each BS and RS and update the dual cost as Eq.(7) descres. It s clearly shown from Fg. that the proposed dstruted power control scheme can quckly converge to the fnal power allocaton result wthout much fluctuaton. Therefore the soluton process can e operated dstrutvely to guarantee the scheme scalalty. Snce the hgh SIR assumptons and dual decomposton are utlzed for the soluton of power control whch leads to suoptmal power allocaton result, the throughput loss of the smplfcatons are nvestgated here. Also the two-cell scenaro s adopted and the throughput of one partcular user n cell s oserved. The throughput otaned y IARA s compared wth 9 3

8 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. Power (W) 6 8 BS RS RS RS3 Per user thoughput (ps) IARA Optmal Iteratons 3 Per lnk SIR (db) Fg. 3: Iteraton convergence of the power control scheme n Sec.IV-B. Fg. : Comparson of IARA and optmal result. Approxmated SIR (db) 3 3 Approxmate SIR Orgnal SIR 3 Per lnk SIR (db) Fg. : Approxmaton of rate expresson of Eq.(3) the actual optmal result (otaned y exhaustve power search) n Fg.. The average SIR of the vares from db to db and the performance gaps wth dfferent SIR are oserved. We regulate the channel fadng of all of the lnks satsfy the Posson dstruton wth the mean value as the gven SIR. We can notce from the results n Fg. that the performance gap ncreases slghtly as the mean SIR decreases. That s manly ecause the omsson of the tem n the left sde of the rate expresson s accurate only at the hgh SIR regon. But the gap s always acceptale wth dfferent SIR regons. What s more, n the scenaro IARA has overlook the nterference that s lower than P th. It s oserved from Fg. that such smplfcaton, whch guarantees the power optmzaton to e done locally, does not lead to much performance degradaton. After that, the mpact of network densty on the aggregate throughput of IARA s compared wth the scheme n [7] as well as the greedy method, whch only ams to optmze the local throughput. We concentrate on the 9-cell scenaro. The cell radus ranges from to 9m. The per-user throughput s represented n Fg.6. It s llustrated from the fgure that oth IARA and the scheme n [7] sgnfcantly outperform the greedy scheme. That s ecause the mult-cell resource allocaton s conducted to allevate the nterference. Results also show that, as the cell densty ncreases, the throughput gan of IARA goes up. When the radus s m, the performance gan s over.3ps per S compared wth the scheme n [7], whle the gan s less than.ps per S when the radus ncreases to 9m. That s manly ecause the scheme n [7] consders the nterference as constant. Actually, as network densty ncreases, the varaton of the nterference ecomes more severe. Our scheme can capture such fluctuaton whle the scheme n [7] cannot. It s also oserved that, the system performance growng speed ncreases as the network densty rases. That s ecause the RS deployment ntroduces more cooperatve gan as the network densty ncreases, whch allevates the advantages of relay transmsson. Snce the specfc RSs are utlzed n the network, the performance gan of edge Ss, whch s the system performance ottleneck, s nvestgated to show the advantages of the delcated RS deployment. In ths scenaro, the last users n Fg.7, whch are nearer to the BS, adopt the drect mode; The Ss whch are at the edge of the BS adopt the relay mode,.e., the frst 3 users n Fg.7. It s clearly oserved from Fg.7 that the average throughput of edge Ss whch adopt the relay mode are enhanced more sgnfcantly than that of nearer Ss. That s generally due to the sgnalng amplfcaton y the RSs as well as ntellgent nterference allevaton. VI. CONCLUSION In ths paper a low-complexty ont sucarrer allocaton, schedulng and power control scheme was proposed for OFDA-ased relay networks y consderng the potental

9 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety suect matter experts for pulcaton n the IEEE INFOCO proceedngs Ths paper was presented as part of the man Techncal Program at IEEE INFOCO. Per user throughput (ps) Cell radus (m) IARA Scheme n [6] Greedy Fg. 6: The mpact of cell densty on system performance. Throughput (ps) IARA Greedy ole staton ID Fg. 7: Comparson of edge Ss and near Ss. nterference among multple neghorng cells. Snce the orgnal ont resource allocaton prolem s NP hard, n our scheme the allocaton process was decoupled nto two stages to reduce the complexty. In the frst stage, the sucarrer and schedulng was conducted y local search. In the second stage, power allocaton was operated for all of the allocated sucarrers. By leveragng geometrc programmng, we prove that optmal power control can e acheved under hgh SIR regon. Smulaton results show that y utlzng the proposed ont resource allocaton scheme, the aggregated throughput of OFDA-ased relay networks can e mproved y over % compared wth the prevous peer work whle the complexty of the scheme s much lower than the centralzed schemes. HKUST ont la, the Natonal Basc Research Program of Chna (973 Program: 7CB367), NSFC-RGC ont program(673663) and the Internatonal ST Cooperaton Program of Chna (ISCP) (No. 8DFA). REFERENCES [] 3GPP TS 36. V8.. 3rd Generaton Prtnershp Proect; Techncal Specfcaton Group Rado Access Network; Evolved Unversal Terrestral Rado Access (E-UTRA); Physcal Channels and odulaton (Release 8), arch 8 [] IEEE standard for local and metropoltan area networks part 6: Ar nterface for fxed roadand wreless access systems, IEEE Std 8.6-, pp. C87, [3] A. Agustn and J. Vdal, Amplfy-and-Forward Cooperaton under Interference-Lmted Spatal Reuse of the Relay Slot,IEEE Transactons on Wreless Communcatons, vol. 7, no., pp. 9-6 ay 8. [] S. De, V. hatre and V. Ramayan, WAX relay networks: opportunstc schedulng to explot multuser dversty and frequency selectvty, n Proc. AC ocom 8, Sept. 8. []. Kodalam and T. Nandagopal, On explotng dversty and spatal reuse n relay-enaled wreless networks, Proc. AC ohoc 8, ay 8. [6] Y. Wu and Z. Nu, Explotng Cooperatve Dversty and Spatal Reuse n ulthop Cellular Networks, Proc. IEEE ICC 9, June 9. [7]. Pschella and J. Belfore, Power control n dstruted cooperatve OFDA cellular networks, IEEE Transactons on Wreless Communcatons, vol. 7, no., pp. 9-96, ay 8. [8] S. Km, X. Wang, and. adhan, Optmal Resource Allocaton n ult-hop OFDA Wreless Networks wth Cooperatve Relay, IEEE Journal on Selected Areas n Communcatons, vol.,no., Fe. 7. [9] B. Can,. Portalsk, H. Lereton, S. Frattas and H. Suraweera, Implementaton Issues for OFD-Based ulthop Cellular Networks, IEEE Communcatons agazne, Sept. 7. [] J. Ca, X. Shen, Jon W. ark and A. Alfa, Sem-Dstruted User Relayng Algorthm for Amplfy-and-Forward Wreless Relay Networks, IEEE Transactons on Wreless Communcatons, vol. 7, no., pp , Aprl 8. []. Awad and X. Shen, OFDA Based Two-hop Cooperatve Relay Network Resources Allocaton, Proc. IEEE ICC 8, June 8. [] B. Gu and L. Cmn, Resource allocaton algorthms for multuser cooperatve OFDA systems wth suchannel permutaton, Proc. of IEEE CISS, Aprl 8. [3] G. L and H. Lu, Resource allocaton for OFDA relay networks wth farness constrants, IEEE Journal on Selected Areas n Communcatons, vol., no., pp. 6-69, Nov. 6. [] H. Holma and A. Toskala, WCDA for UTS. Second Edton, Hooken, NJ: Wley,. []. Chang, C. Tan, D. Palomar, D. O Nell and D. Julan, Power control y geometrc programmng, IEEE Transactons on Wreless Communcatons, vol. 6, no. 7, pp. 6-6, July 7. [6] A Bletsas, A Khst, DP Reed and A. Lppman, A Smple Cooperatve Dversty ethod Based on Network Path Selecton,n IEEE Journal on Selected Areas n Communcatons, vol., no. 3, arch 6. [7] E. Lo and K.B. Letaef, Optmzng Downlnk Throughput wth User Cooperaton and Schedulng n Adaptve Cellular Networks, Proc. IEEE WCNC 7, ay 7. [8] K.J. Ray Lu, A. Sadek, W. Su and A. Kwasnsk, Cooperatve Communcatons and Networkng, Camrdge Unv. Express, 9. [9] S. Boyd and L. Vandenerghe, Convex Optmzaton, U.K.: Camrdge Unv. Press, ACKNOWLEDGENT The research was supported n part y Hong Kong RGC Grants HKUST 639, 68, N HKUST69/7, Huawe-