Clustering Based Fractional Frequency Reuse and Fair Resource Allocation in Multi-cell Networks

Transcription

1 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety subject matter experts for publcaton n the IEEE ICC 21 proceedngs Clusterng Based Fractonal Frequency Reuse and Far Resource Allocaton n Mult-cell Networks Wehuang Fu, Zhfeng Tao, Jnyun Zhang, and Dharma P. Agrawal Center for Dstrbuted and Moble Computng Department of Computer Scence, Unversty of Cncnnat, Cncnnat, OH 45221, USA Mtsubsh Electrc Research Laboratores (MERL), Cambrdge, MA 2139, USA Emal: fuwg@cs.uc.edu, tao@merl.com, jzhang@merl.com, and dpa@cs.uc.edu Abstract Fractonal frequency reuse (FFR), usng dfferent frequency reuse factors for cell center and edge regons, s able to effectvely mprove spectrum effcency n mult-cell OFDMA networks. However, optmal performance s hard to acheve n practce as the effcency of resource allocaton could drop drastcally due to the constrant from the frequency parttons formed by FFR. Snce the rado resource s pre-parttoned for cell edge and center, far resource allocaton n a cell s also dffcult to mplement. Conventonal frequency partton adjustment ether has hgh complexty due to global optmzaton or suffers from heavy performance degradaton due to absence of effectve control on nter-cell nterference (ICI). To solve ths ssue, we create models for analyzng geographcal dstrbuton of nterference n mult-cell networks. Based on the observed non-unform dstrbuted ICI, we redefne the zones for fractonal reuse and propose clusterng based FFR, whch offers resource allocaton hgher flexblty and better farness wth addtonal spatal dmenson. Extensve smulaton has been performed to valdate practcalty and effectveness of our proposed scheme. I. INTRODUCTION Frequency reuse s utlzed by mult-cell systems to mprove system throughput, whch nevtably ntroduces nter-cell nterference (ICI) and compromses qualty of servce. As the frequency resource s spatally reused, moble statons (MSs) may be nterfered by addtve sgnals from multple neghborng cells operatng on the same rado at the same tme. To mtgate ICI, prevous mult-cell systems dvdes frequency band nto multple orthogonal parttons and dentcal partton s only reused at cells wth certan dstance away. The number of parttons, defned as frequency reuse factor, can be used to dentfy the space dstrbuton of frequency resource. Hgher frequency reuse factor can reduce ICI sgnfcantly; however, t also greatly decreases the spectrum effcency due to less frequency resource avalable at each cell. The emergence of a mult-user verson of orthogonal frequency-dvson multplexng (OFDM),.e., OFDMA, facltates mplementaton of a more sophstcated technology fractonal frequency reuse (FFR) [1][2][3][4]. OFDMA technology, whch s wdely adopted n most next generaton mult-cell cellular systems, such as 3GPP Long Term Evoluton (LTE) [5] and IEEE 82.16, enables the use and allocaton of resource n both frequency and tme domans. Wehuang Fu worked on ths study whle vstng Mtsubsh Electrc Research Lab, Cambrdge, MA. Although system performance revenue n utlzng FFR s theoretcally observable, challengng problems are also presented n actual practce. The flexblty of resource allocaton wll be serously constraned by the pattern of frequency parttons, whch may lead to performance degradaton n the dstncton of traffc load, servce requrement, number of MSs, etc. The problem also presents n performng far resource allocaton n mult-cell networks. Conventonal FFR s not scalable n solvng dfferentated frequency parttons and resource allocaton. The adjustment of frequency partton n a cell greatly affect ts neghborng cells. As cells are back to back to provde a seamless coverage, the effect of frequency parttons n a cell can be observed on other cells. Although global optmzaton has been proposed for FFR to adjust the parttoned amount and frequency reuse factors, to gve hgher spectrum effcency, t s rather hard to mplement due to scalable problems n nformaton exchange overhead, computaton complexty, etc. In ths paper, we analyze the geographcal dstrbuton of nterference n mult-cell networks. Based on the observed nonunform dstrbuted ICI, we reform the zones for fractonal reuse and propose a clusterng based FFR, whch offers hgher flexblty n resource allocaton wth addtonal spatal adjustment. The performance gans are manly due to the followng advantages: Intra-cluster resource allocaton jontly mtgates ICI wth fractonal resource reuse and provdes proportonal far schedulng (PFS). Dfferent from a cell based allocaton that has a large cell regon sufferng from ICI, our proposed clusterng based allocaton has much smaller regon havng potental nter-cluster nterference. Thus, much less nformaton exchange s needed n mtgatng nter-cluster nterference. A sgnfcant performance enhancement from proposed scheme s observed by the smulaton results as compared to cell based FFR and cell based FFR-PFS. II. PROBLEM FORMULATION AND MODELING We consder an OFDMA network consstng of K BSs and N MSs, where the set of BSs s denoted by K = {1,...,K}. The locatons of BSs are denoted by S 1,...,S K, respectvely. BS provdes data servce for MSs n the coverage area, usually a crcle centered at the coordnates of the BS wth radus R T n a two-dmensonal area. Ths s an area where MS can correctly decode downlnk frame preamble f no co-channel /1/$ IEEE

2 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety subject matter experts for publcaton n the IEEE ICC 21 proceedngs Fg. 1: Mult-cell and FFR nterference exsts. To cover a large regon, multple BSs are placed as hexagon cells. Inter-BS dstance, also called ste to ste dstance, s denoted by D s. Assume K BSs are connected to each other by the wred backhaul network and form a multcell system to provde servce for the MSs n the entre coverage area. A non-overlappng coverage of BS s a hexagonal area, where the nradus of hexagonal, denoted by R HI,s Ds 2, and crcumradus, denoted by R HC,s Ds 3. We consder such a hexagonal area as a cell formed by a BS, and a cell formed by BS u s called cell u. The example for a part of the network s shown n Fg. 1. MS may be covered by multple BSs. Three stuatons of the MS locatons are llustrated n the fgure. MS 1 located n the area where t s only n the coverage of BS 1.The MS has strong sgnals from BS 1 and weak nterference from other neghborng BSs. In the second stuaton, MS s n the area overlapped by two cell BSs, so t s able to receve the broadcast messages from both BSs and acknowledge both BS dentfcatons (IDs). For example, MS 2 n Fg. 1 can receve the messages from BS 1 and BS 3. MS s only allowed to regster at one of the BSs, whch s called the servng (or anchor) BS, and the MS only has the data exchange wth the servng BS. We assume MS n the network takes the BS wth the mnmum Eucldean dstance as the servng BS, and executes the handover process to connect to the target BS when the target BS becomes closer than the servng BS. Smlarly, the MS located n the area overlapped by the coverage area of three BSs s able to acknowledge three BS IDs lke MS 3 n the fgure. Every MS wll record the BS IDs from whch t receves perodc broadcastng messages. The set of BS IDs s called dversty set, whch wll be reported to servng BS by MSs perodcally. To mprove the spectrum effcency, BS cell s dvded nto three sectors wth three separated antenna sets. The antenna pattern for each sector refers to [6]. OFDMA frames n the system are synchronzed and have the same tme dvson duplex (TDD) for downlnk (DL) and uplnk (UL). TDD frame s called DL frame and UL frame respectvely. Although, we dscuss the resource allocaton for DL frames n ths paper, the method s equally applcable to UL frames. DL frame can be dvded nto multple resource blocks (RBs) [1] whch are used as the basc unts for resource allocaton. Each RB s comprsed of multple subcarrers and OFDMA symbols. Logcally, the RBs avalable for data transmsson constructs a two-dmensonal plane. Let M f ndcate the number of RBs spannng over frequency doman and M t denote the number of RBs spannng over tme doman. RB n a DL frame s dentfed by frequency and tme doman par (f,t) where f {1, 2,..., M f } and t {1, 2,..., M t }.RBalso can be ndcated by ndex, where =(f 1)M f + t. Correct detecton of sgnal depends on sgnal to nterference and nose rato (SINR) at recever [7]. The SINR for MS located at pont s n the cell of BS u n RB can be expressed by: ξ (u) (s) = v U u P (u) S u s α, (1) P (v) S v s α + N o where U represents the set of BSs usng RB, P (v) denotes the transmt power of BS v over RB, α (α 2) denotes the sgnal attenuaton factor, and N o shows the thermal nose power. In Eq. (1), f the dstance between MS and BS u s fxed, ξ (u) (s) s manly affected by S v s, the dstance between the MS and other nterferng BSs. If s s on the cell margn, S u s wll be close to S v s. Whle neghborng cells are usng dentcal RBs, ξ (u) (s) n Eq. (1) could be lower than the threshold of correct recepton, whch means the sgnals of neghborng cells are strongly nterferng to each other. If dentcal RBs are used by far enough located BSs, the raton of S u s and S v s (s) can be (s) becomes larger f S u s s reduced. FFR takes advantage of non-unform SINR n a cell and dvdes a cell nto cell center and cell edge zones, usng dfferent frequency reuse factors. As shown n Fg. 1, the frequency band s parttoned, where cell center zone can use overlapped frequency sub-band F c wth transmt power P c1. The rest of frequency band s dvded nto three parttons for the edge zone of three sectors, whch are F e1, F e2, and F e3, respectvely. Wth such a method, the frequency sub-band F c has reuse factor one, whle the rest frequency partton s reuse factor s three. RBs belongng to dfferent parttons can be assgned to MSs n the correspondng zones. To facltate MS zone determnaton, t can utlze MS s dversty set. MS wth more than one BS n the dversty set s consdered n edge zone; otherwse t s n center zone. Resource allocaton for MSs n dfferent zones s subjected to the allocated frequency partton. For example, for MSs located n the edge zone of sector 1 n BS 3, the resource allocaton s only allowed to be wthn F c1. The szes of frequency parttons hghly lmts the flexblty of resource allocaton. It s dffcult to adjust the sze of frequency parttons as dvsons of frequency partton s strongly related. can be large, whch ndcates weak ICI and ξ (u) hgher than the threshold. ξ (u) III. NETWORK CLUSTERING ICI s not unformly dstrbuted n a cell. For the MSs located at cell edge zone, some are manly affected by ICI from one BS, and the rest wll have strong ICI from two neghborng BSs. In contrast, the area around BS, whch s the cell center zone, has weak ICI. As shown n Fg. 2, the lght color area at the cell edge of BS 1 denotes the area where ICI manly

3 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety subject matter experts for publcaton n the IEEE ICC 21 proceedngs from one neghborng BS, and the dark color area n BS 1 cell denotes the area where ICI manly from two neghborng BSs. Dfferent from conventonal FFR, we further consder the fundamental reasons and utlze non-unformly dstrbuted ICI for performance mprovement. Our method s based on the cluster formed by adjacent sectors, where most ICI s present. Clusterng based FFR reformats the zones to elaborately utlze non-unform SINR and enable far resource schedulng. A general cluster s formed by adjacent sectors of adjacent cells. Cluster on the boundary regon of network could be formed by two sectors snce only two sectors are adjacent. Some cluster may only have one sector, whch does not have to exchange sector nformaton through the backhaul. In the rest of the paper, for dscussons, we take the cluster havng three sectors as a general cluster n the network. The clusters on the boundary are the specal case of the typcal cluster. The methods and algorthms are also applcable to these clusters. As shown n Fg. 2, cell edge zones from three adjacent sectors are combned together. The area where ICI s manly from two BSs, are separated from tradtonal cell edge zone and s defned as cluster corner zone. The reason behnd ths cluster formaton s that nter-cell nterference s most serous n the ntersecton of three adjacent cells. The BSs that wll cause each other s ICI at the cell edge zones (except for the cluster corner zone) are ncluded n the cluster. By jontly allocatng resource n the formed cluster, ICI can be effectvely avoded or mtgated. The resource allocaton of MSs are related to ther zones. Determnaton of the zone of an MS depends on ts dversty set. If the dversty set only has the Identfcaton (ID) of the servng BS, t ndcates that the MS s n the cell center zone. If the dversty set has two BS IDs, t ndcates the MS s n the cell edge zone. If the dversty set has three BS IDs and these are ncluded n the same cluster, t ndcates the MS n the cell edge zone. More specfcally, t s located at the ntersecton of three sectors. If the dversty set have three BS IDs and any of them s not n the cluster, t ndcates that the MS s at the cluster corner zone. As llustrated n Fg. 2, the corner zone frequency can be fxed as smlar to conventonal FFR, ths reduces the overhead n adjustng nter-cluster nterference. Snce the sze of cluster corner zone s very small, the resource for that wll not be too much. The resource allocaton for cell edge zone can be adjusted wthn a large regon, whch offers hgher flexble than conventonal cell-based FFR. The allowed adjustment n conventonal FFR for edge zone s between and F e. Wth clusterng based FFR, t can be from to 3F e F corner. Snce a cell s dvded nto three sectors, the management of resource at BS can be dvded nto three ndependent parts for three sectors. For the sectors n a cluster, the BS of the sector leaves the resource management of the sector to the cluster. For a BS nvolved n three dfferent clusters, the three sectors belong to three dfferent clusters. Every cluster has a cluster head n charge of the resource management and allocaton. Clusters are formed at the ntaton of the network and are Fg. 2: Proposed clusterng based FFR updated when BS avalablty changes. For example, when a new BS jons the network, clusters wll be reformed for the new BS. Or BS leaves the network, clusters also wll be reformed agan. The reformaton s dstrbuted and only happens at the sectors near the place where any BS exstng changes. After performng the resource allocaton, the nformaton wll be sent to the correspondng BS through the backhaul. The BSs then can perform DL transmssons accordng to the schedulng. The correspondng transmt power of RB s determned by the zone of MSs, and the data rate are determned by the obtaned channel state nformaton (CSI). IV. RESOURCE ALLOCATION Our resource allocaton s to cooperatvely allocate RBs of three DL frames n a cluster, as shown n Fg. 3. Dfferent from RB resource n a square two-dmensonal plane, RB s a resource cube n three-dmensonal space by addng the dmenson of the frame. RB s dentfed by (f,t,s) n three-dmensonal space, where s s the ndex of DL frame, s {1, 2, 3}. Three resource cubes wth the same s construct a resource cubod, whch s dentfed by (f,t). Smlar to prevous dscussons, a resource cubod also can be dentfed by ndex, where =(f 1)M f + t. So, when we menton the th resource cubod latter, we mean the resource cubod (f,t). The allocaton of the three DL frames can be jontly optmzed by the cluster head. The allocaton s subjected to certan constrants. When a resource cubod s used for allocaton of MS at the cell center zone, we mplement frequency reuse factor one as used by FFR to acheve a hgher spectrum reuse factor. Snce the SINR are relatvely hgh for MSs n ths zone, the system s to take advantage of hgh reuse factor so as to mprove the throughput. So, three RBs n a resource cubod wll be allocated to MSs n the cell center zone of the correspondng sectors as shown n Fg. 3(a). For example, for the kth resource cubod, RB (, 1) for the MS s allocated n the cell center zone of sector 1, RB (, 2) s for the MS n the cell zone of sector 2, and so on. So, a resource cubod can support up to three MSs n dfferent cell center zones. Let a bnary varable A (u) (, s) denote the allocaton of RB of sector s. A (u) (, s) 3, u Z c (s), (2) u s s

4 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety subject matter experts for publcaton n the IEEE ICC 21 proceedngs TABLE I: Dscrete AMC Modulaton BPSK 4-QAM 16-QAM 64-QAM 256-QAM r u b/sym I u (db) Fg. 3: Resource allocaton for zones where Z c (s) denote the set of MSs n the cell center zone of sector s. When a resource cubod s utlzed for allocaton of MS at the cell edge zone, the prmary purpose s to mtgate nter-cell nterference. So, the whole resource cubod s only allocated for one MS, nstead of three MSs, as llustrated n Fg. 3(b). The allocaton constrant s: A (u) (, s) 1, u Z e (s), (3) u s s where Z e (s) denote the set of MSs n the cell edge zone of sector s. Due to RB can only be used once by a sector, an addtonal constrant s: A (u) (, s) 1, u Z c (s) Z e (s). (4) u The allocaton for cluster corner zone s to use the fractonal RB reuse to avod complex nter-cluster nformaton exchange. The resource blocks used for the corner zones are pre-planed for the reuse at dfferent clusters, as llustrated by Fg. 3(c). For example, f the resource cubod s used by a BS for the cluster corner zones, t wll not be used by other BS n the same cluster, but can be reused only by the cluster corner zones of BSs at a reuse dstance. Gven data RBs of a DL frame, we need an approprate schedulng scheme to allocate resource cubods. To farly perform ntra-cluster resource block allocaton, our scheme mplements proportonal far scheduler (PFS), whch means each MS has proportonal data rate. To maxmze the overall data rate, scheduler enables the MSs havng bad channel condton to have more RBs so as to have smlar data rate as those havng good channel condtons. Also, dfferent from a scheduler consderng far number of RBs, t takes the modulaton and codng rate nto consderaton and provdes proportonal farness n terms of data rate. To allocate RBs for the MSs n the cluster, the scheduler scans the resource cubods by sequence of {1, 2,...,k,...,M RB }, where M RB s less than M f M t. Snce certan RBs are reserved for the cluster corner zone, the RBs from M RB +1 to M f M t areusedfor the cluster corner zone. The scheduler selects MS u wth the mnmum metrc ρ u from all MSs n the cell center and cell edge zones to allocate the resource and consder two cases: cell center zone MS and cell edge zone MS, whch are dscussed earler. Metrc ρ u s computed by: ρ u = A(u) (, s)r u R u, u Z(s), (5) where r u s determned by effectve sgnal to nose rato (SNR) I u of MS u, Au (, s) denote the number of RBs allocated to MS u, and R u s the data rate of MS u of the last frame. Cluster head gathers CSI nformaton and selects approprate AMC for MSs. The selecton of r u for MS u follows Tab. I. On the other sde, data rate R u s updated per DL frame: R u = α A (u) (, s)r u +(1 α)r u, (6) where R u s the data rate of current DL frame, R u s the data rate of the prevous DL frame, and α s the decay factor that controls the nfluence of past data rate and currently allocated data rates. The objectve s to approach a long-term farness. V. PERFORMANCE EVALUATION In ths secton, we nvestgate the performance of the proposed schemes. The confguraton of smulaton follows the suggestons n the IEEE 82.16m evaluaton methodology document [6]. We frstly have a look at the effectve SINR of MSs. Because the SINR wll be dfferent from the RBs allocated to an MS, we measure the effectve SINR for MSs, whch s computed from SINR of each allocated RB. Fg. 4 and Fg. 5 show the cumulatve dstrbuton functon (CDF) of the MS effectve SINR n the network wth reuse-1 allocaton and clusterng based allocaton. Obvously, the SINR of clusterng based allocaton s hgher than that of reuse-1 allocaton. Ths s because the MS n the edge zone of reuse- 1 wll suffer from hgh nterference due to the resource reuse of the neghborng cells. Some MSs cannot receve any data due to low SINR. If gven a threshold 2dB as the mnmum SINR for communcaton, reuse-1 allocaton wll have about 7 percent SINR below the threshold. However, n clusterng based allocaton, the MSs n the edge zone s mtgated by a cooperatve allocaton between the sectors. The SINR wll be hgher than the threshold so that all MSs n the network can effectvely do the communcaton. In addton, the span of clusterng based allocaton s narrower than that of reuse- 1, whch means the SINR n clusterng based allocaton has smaller devaton. In reuse-1, the SINR hghly reles on the geographcal locaton of MS. MS closer to the BS would have much hgher SINR. Clusterng based allocaton mproves SINR of those MSs n the edge zone, so the dstrbuton of SINR s closer to the average SINR. The trends can be nvestgated both n lght traffc load stuaton of Fg. 4 and heavy traffc load stuaton of Fg. 5. Fg. 6 shows the throughput per cell of three allocaton methods: cell based FFR, cell based FFR-PFS, and clusterng

5 Ths full text paper was peer revewed at the drecton of IEEE Communcatons Socety subject matter experts for publcaton n the IEEE ICC 21 proceedngs 1.9 Reuse Reuse x CDF CDF Throughput/Cell Effectve SINR (db) Effectve SINR (db) PFS MS/Cell Fg. 4: Effectve SINR (lght traffc) Fg. 5: Effectve SINR (heavy traffc) Fg. 6: Throughput per cell 15 x 16 PFS.35.3 PFS.8.7 PFS Throughput/MS 1 5 Percentage Percentage MS/Cell Data Rate x Data Rate x 1 6 Fg. 7: Throughput per MS Fg. 8: Data rate (lght traffc) Fg. 9: Data rate (heavy traffc) based PFS, where throughput s computed by Shannon capacty equatons. Clusterng based PFS has hgher throughput per cell than these of cell based FFR and cell based FFR-PFS. Due to frequency resource reuse wth fxed dstance, FFR s not able to fully reuse all resources. Wth dfferent traffc load, clusterng based PFS has almost consstently better throughput as the throughput of cell based FFR and cell based FFR-PFS are changed wth traffc load. For lght traffc load, clusterng based PFS has much hgher throughput than that of cell based FFR and cell based FFR-PFS. Ths s one of the mportant advantages of clusterng based allocaton, because t s able to utlze the resource not used by neghborng sectors. When the traffc load s lght, the MSs are not dstrbuted so unformly. The devaton of the number of MSs n a sector and/or the number of MSs n the center and edge zones s large. The resource wll be wasted n a cell based FFR and cell based FFR-PFS snce they are not able to cooperatvely allocate the resources. Clusterng based allocaton can fully utlze the dle resource of neghborng sectors for heavy loaded sector. From the fgure, cell-based FFR has hgher throughput than cell-based FFR-PFS. Because cell-based FFR-PFS s to reach the far rate, the overall throughput s not maxmzed. Fg. 7 shows the throughput per MS. It ndcates a smlar trend as n Fg. 6. Clusterng based PFS stll has the hghest throughput and the throughput of cell based FFR s hgher than that of cell based FFR-PFS. But, they all approach a lmt when the traffc load ncreases. Fgs. 8 and 9 respectvely show the hstograms of the data rate wth lght traffc load and heavy traffc load. In Fg. 9, t s obvous that clusterng based FPS farly concentrates around certan data rate. More than 7 percent of data rate are scaled n a narrow data rate scope n Fg. 9. Cell based FFR and cell based FFR-PFS have a data rate wth much wder scope. Whle the traffc load s lght, the data rate s more wdely dstrbuted. Snce some RBs cannot be used for cell edge zone or other cell center zone n the same cluster, the allocaton of resource cubods for the cell center zone wll allocate all the RBs to MSs so as to maxmze the throughput. So, some MSs wll have hgher throughput. When the traffc load s lght, devaton of MS geographc locaton s relatvely hgh, makng the effect more obvous. VI. CONCLUSION Based on the observed non-unform dstrbuted ICI, our proposed method wth clusterng of adjacent sectors has better performance n varous aspects. Consderng the advantages n scalablty, throughput, and farness, our scheme can be utlzed by current mult-cell OFDMA network standards, such as the IEEE or the 3GPP LTE, so that spectrum effcency and flexblty can be mproved. REFERENCES [1] The draft IEEE 82.16m system descrpton document, n IEEE Broadband Wreless Access Workng Group. [2] R. Gulano and C. Mont, WMAX fractonal frequency reuse for rural envronments, n Wreless Communcatons, June 28. [3] Y. Zhou, Smulaton study of fractonal frequency reuse for moble WMAX, n VTC, 28. [4] Comment for nterference mtgaton usng fractonal frequency reuse, n IEEE c82.16m-8/1383. [5] M. C. Necker, Local nterference coordnaton n cellular OFDMA networks, n VTC, 27. [6] Project 82.16m evaluaton methodology document (EMD), n IEEE Broadband Wreless Access Workng Group. [7] W. Fu and D. P. Agrawal, Effectve rado parttonng and effcent queue management schemes n a wreless mesh network, n Globecom, 28.