01 7th Internatonal ICST Conference on Communcatons and Networng n Chna (CHINACOM) Performance Evaluaton of Layer Adaptve Mult-User Schedulng n LTE-A Downln L Zhang, Sov Peter, Chunye Wang Chna Industral Envronment, Chef Technology Offce, Noa Semens Networs, Bejng, Chna l.7.zhang@nsn.com; peter.sov@nsn.com; chunye.wang@nsn.com Abstract Mult-user MIMO s an mportant technque to utlze the spatal degree of freedom brought by multple antennas at the transmtter. To be employed n the downln of practcal systems, some sub-optmal algorthms have to be developed to handle the user selecton and precodng wth lmted feedbac on channel state nformaton and reasonable computatonal complexty at a base staton. In ths paper we evaluate the performance of a specfc mult-user schedulng algorthm, whch selects users one by one wth greedy search and results n a scheduled user set wth adaptve sze. The performance s compared wth tradtonal two-user schedulng algorthm n sngle-cell, and we also show the performance of Coordnated Mult-Pont transmsson wth the layer adaptve algorthm by applyng t to users from multple cells. The system level smulaton s based on 3GPP LTE-A framewor, and the evaluaton can be used as gudance for concrete scheduler desgn. Key Words mult-user MIMO, user schedulng, Coordnated Mult-Pont transmsson, LTE-A I. INTRODUCTION Mult-user MIMO has been a hot topc for both research and ndustral lteratures n recent years. The basc dea s to transmt multple data streams to dfferent users (recevers) on the same tme-frequency resource. As an mportant technque to utlze the spatal degree of freedom brought by multple antennas at the transmtter, mult-user MIMO has some advantages over sngle-user MIMO [1] such as mmunty to propagaton lmtatons (e.g. per user channel matrx does not need to be full ran) and reduced requrements on the recever capablty (e.g. mult-layer processng s not requred). The nvestgatons have been focused on dfferent aspects of mult-user MIMO. The nformaton theoretcal channel capacty has been extensvely studed, along wth optmal precodng schemes []. Due to the prohbtve complexty n the optmal drty paper codng (DPC), some low complexty precodng schemes have been studed, such as zero-forcng (ZF) and bloc-dagonalzaton (BD). The number of scheduled users s drectly lmted by the number of antennas at the transmtter wth lnear precodng schemes, and the problem of how to select scheduled users arses. The scenaro of nterest of ths paper s downln of a cellular system, and here the optmal set of scheduled users can be defned wth a specfc optmzaton target and derved by exhaustve search over all possble combnatons. However, the complexty s huge wth medum to large number of canddates n a cell, so some sub-optmal user schedulng algorthms are desrable for practcal base staton mplementaton. To ths end, some greedy search algorthms have been proposed. In [3] the frst user s selected through channel gan, and followng users are added one by one untl an arbtrary number s reached. At each teraton, the user whch maxmzes sum rate (f t s added) s selected wth ZF precodng taen nto account. Same process s proposed n [4][5], but the lower bound of the broadcast channel capacty and the sum of channel gans wth null space successve projecton are used at each teraton as optmzaton target, respectvely. Currently coordnated mult-pont transmsson (CoMP) s beng studed and standardzed n 3GPP for LTE-A [6]. One transmsson scheme of CoMP s coherent jont transmsson, where data stream for a sngle user s coherently precoded and transmtted from antennas at multple base statons. To verfy the gan of mult-cell coordnaton, the performance of sngleuser MIMO, mult-user MIMO and CoMP are often compared through system smulatons [7][8], snce t s very dffcult to provde analytcal evaluatons wth realstc envronments. A smple and effcent user schedulng algorthm, whch can be vewed as varant of those n [3-5], was frstly proposed n [9]. In ths paper, we provde more detaled nvestgatons for the algorthm; compare ts performance wth tradtonal two-user schedulng algorthm n sngle-cell under dfferent assumptons. Smlar as [9], we also show the performance of CoMP wth ths algorthm but appled to users from multple cells. Snce the throughput gan from schedulng more than two users depends a lot on the assumptons, the evaluaton n ths paper can be used as gudance for concrete scheduler desgn for both sngle-cell and mult-cell operaton. The rest of the paper s organzed as follows. System model and smulaton setup are dscussed n Secton II. Secton III descrbes the layer adaptve user schedulng algorthm, and also the tradtonal two-user schedulng algorthm for comparson. In secton IV performance evaluaton s provded based on system level smulatons. Fnally, Secton V concludes the paper. 787 978-1-4673-699-5/1/$31.00 01 IEEE
II. SYSTEM MODEL AND SIMULATION SETUP Ths paper s concerned wth LTE-A downln, where data for multple users are transmtted from base statons. A. Sngle-cell scenaro In sngle-cell scenaro, each base staton s assgned to a cell to handle the users n the cell coverage area, as shown n Fgure 1. The basc multplexng scheme n LTE-A s orthogonal frequency dvson multplexng (OFDM), whch means users n a same cell wll be allocated wth dfferent resource blocs (RBs). Wth mult-user MIMO, a sngle RB can be shared by multple spatally multplexed users. In ths paper we only consder one data stream for each user,.e. there s no spatal multplexng of sngle-user MIMO. Fgure 1 Sngle-cell scenaro consdered n LTE-A downln Denote N as number of schedulable users n -th cell, and Ω as the set of scheduled users n the same cell on a specfc RB. For the sae of smplcty, RB ndex s not present n the system model n ths secton as rado resource management s same across RBs. The receved sgnal at user n -th cell, can be modeled as r = H w x + H w x + H w x + n ( 1 ),, l l j, l l l Ω, l j, l Ω j, where H j, s the channel matrx between user and base staton j, w l and x l are the precodng vector and data symbol of user l, n s the nose at the recever of user. In ths paper the precodng matrx n -th cell s calculated by ZF over precodng vectors from spatally multplexed users: H ( ) 1 W =V V V V = v,v,...,v where S s the sze of Ω. 1 S ( ) In realstc world, there wll be channel estmaton error on SRS. Besdes, some other mperfectons are also consdered n ths paper, namely, sngle-antenna (ssrs) and narrow band (nsrs). ssrs means users wll transmt SRS from only one antenna, although equpped wth multple antennas for recever; as a result, only one row of the full channel matrx wll be nown at the base staton. nsrs means users wll transmt SRS n part of the system bandwdth, and effectvely the channel nformaton on a specfc RB wll be more lely to be outdated. In frequency dvson duplex (FDD) system, base staton has to rely on feedbac from users to derve per-user precodng vector n ( ). In LTE-A, codeboo based method s defned wth precodng matrx ndcator (PMI) feedbac. Gven the receved sgnal n ( 1 ), sgnal-to-nterferencenose-rato (SINR) s calculated for each user as SINR = uh, w j, l l + j, l l + σ n j=, l j, l u H w H w ( 4 ) where u s the combnng vector at user. Assumng perfect nterference nowledge at user recever, u can be derved wth mnmze mean square error (MMSE) prncple as 1 ( σ ) (, ) ( jl, l)( jl, l) ( jl, l)( jl, l) (, )(, ) u = R+ I H w n jl =, jl, H H H R= H w H w + H w H w H w H w ( 5 ) where R s the nterference covarance matrx, wth the frst term from transmsson n -th cell, whch s noted as nteruser nterference (IUI), and the second term accounts for ntercell nterference (ICI). In ths paper we assume perfect nowledge about IUI and ICI are avalable at user recever; however, wth realstc recever, the accurate covarance matrx for ICI part cannot be fully nown, but the power estmaton s usually used for dagonal elements and zero for other elements. B. Mult-cell scenaro In mult-cell scenaro, a sngle base staton s assgned to handle the users n the coverage area of multple cells, as shown n Fgure. Ideally, f perfect channel nowledge s avalable, e.g. n tme dvson duplex (TDD) system where downln channel can be measured from upln soundng reference sgnal (SRS), v s the domnant egen vector of the channel. H H, =U, S, V, ( 3 ) Fgure Mult-cell scenaro consdered n LTE-A downln 788
Wth remote rado head (RRH) deployment or based band poolng technque, a super cell can be formed to handle more users n a larger area; for example, users n one cell wll only receve desred sgnal from antennas at one of the three colocated base statons as n Fgure 1, but n Fgure all the antennas at the cell ste are used transmt desred sgnal for users n the whole coverage area. The specfc scenaro n Fgure s mentoned as ntra-ste CoMP. There are some benefts of such jont processng from mult-user pont of vew, and one aspect s that precodng matrxes whch are calculated separately n sngle-cell scenaro are now combned coherently, so ZF can be performed across more users. Ths wll reduce the nterference experenced by users, as ICI n sngle-cell scenaro are now treated as IUI. III. LAYER ADAPTIVE USER SCHEDULING One mportant aspect of mult-user MIMO n realstc world s the user schedulng,.e. how to determne Ω accordng to the channels states { H,, { 1,,..., N} }. Bascally, there are two factors that affect the selecton, namely the per-user schedulablty and compatblty among users. The former can be measured by the schedulng metrc used n sngle-user MIMO, whle the later can be based on the orthogonalty between the channels of dfferent users. A straghtforward user schedulng algorthm as n Table 1 can be developed, whch always schedules two users on each RB (named as MU-). The prmary user s selected by snglestream proportonal far (PF) metrc, and the secondary user s selected accordng to the product of sngle-stream PF metrc and orthogonalty wth the prmary user. Table 1 A straghtforward schedulng algorthm that always schedules two users on each RB 1) prmary user selecton 1 = arg max ) secondary user selecton H v 1 = arg max 1 1 v 1 The lmtaton of MU- s that two users can be scheduled at maxmum, and there could be potental throughput gan from schedulng more than two users. In partcular, t s mpossble to extend the algorthm n mult-cell scenaro as n Fgure, where degree of freedom n spatal doman wll not be utlzed effcently wth only two users scheduled, as the number of antennas wll be large. Therefore, an algorthm that schedules adaptve number of users s more desrable. Table shows a smple and effcent layer adaptve user schedulng algorthm [9] (named as MU-LA). Arbtrary A users can be scheduled on the same RB, f enough qualfed users can be found. At each teraton, the canddate user set s refreshed so that every qualfed user s orthogonal to all the users n the scheduled user set. If the canddate user set s not empty, the user wth hghest PF metrc s selected and added to the scheduled user set; otherwse, the algorthm stops wth less than A users scheduled. Table A layer adaptve schedulng algorthm that always schedules arbtrary number of users on each RB 1) prmary user selecton 1 = arg max Set ter=0 Set scheduled user set ω = { 1 } Set canddate user set θ = { 1,,..., N} 1 ) secondary users selecton Whle ω < A Increase ter by 1 For θ H vl If > OThreshold, l ω vl θ = θ End f End for If canddate user set θ > 0 = arg max θ Set ω = ω + Else Brea End f End whle The advantage of MU-LA s the adaptve number of scheduled users, whch could be adjusted accordng to the antennas at the base staton. Moreover, t can be easly appled to mult-cell scenaro, by trplng A for ntra-ste CoMP. IV. SIMULATION ANALYSIS In ths secton we wll provde performance evaluaton for MU-LA. The frst step s to nvestgate what could the best value be for O Threshold n Table. Then wth ths optmzed threshold, MU-LA s compared to MU- n sngle-cell scenaro under dfferent assumptons. Fnally, smulaton results are shown for mult-cell scenaro wth MU-LA. A. Smulaton paprameters The system level smulator s bult based on LTE-A downln. Homogenous networ s consdered wth full buffer traffc for all users, whch can provde a benchmar for the evaluaton of spectral effcency of dfferent MIMO technques. Users are unformly dstrbuted n the area of 19 stes/57 cells, and wrap around s enabled to model the out-ter nterference. For each cell, both 4 and 8 (cross-polarzed) transmt antennas are smulated, and both deal and realstc users recever (n terms of nterference nowledge) are consdered. Channel qualty ndcator (CQI) s calculated on common reference sgnals, whch are not precoded. As a result, outer loop ln adaptaton s employed [10] to compensate the msmatch between the measured reference sgnal strength and the power for data transmsson, and t s partcularly useful for MU-LA snce the transmt power at base staton s splt to adaptve number of scheduled users. Detaled smulaton parameters can be found n Table 3 (settngs n bold are default ones). 789
System Parameter Table 3 Smulaton assumptons for sngle-cell scenaro Settng Smulaton scenaro Macro Case 1 wth 19 stes(57 cells) Frequency band 10 MHz bandwdth at GHz carrer frequency Channel model 3GPP SCM UMa 8 azmuth spread (NLOS) / ITU UM Number of users per cell 10 Base staton antenna Columns wth {-45, +45} deg. x-pol antennas confguraton 0.5-wavelength spacng between columns 4/8 antenna elements per cell UE antenna confguraton Rx, x-pol wth {0,90} deg. Precodng scheme SVD/PMI based, wth ZF on top Recever User schedulng SRS confguraton CQI/PMI feedbac Ln adaptaton Ideal SU: sngle-cell PF schedulng wth RA MU-LA: max. /4 users scheduled for 4/8Tx MU-: fxed user schedulng 10 ms perod Ideal: dual antenna, 48-PRB Realstc: sngle antenna, 16-PRB Sub-band report based on CRS ports 6 ms delay, 10 ms perod Based on CQI and 10% frst BLER target B. Optmal value for O Threshold OThreshold wll determne how lely one user could fulfll the orthogonalty condton and can be qualfed as canddate user. As a threshold for the channel correlaton between two users, hgher value means looser requrement, and more users can be scheduled on the same RB. On the other hand, the orthogonalty among scheduled users becomes worse, and ZF would cause more severe power penalty to the desred sgnal. Table 4 shows the dstrbutons for number of scheduled users per RB, wth dfferent OThreshold values, and Fgure 3 shows the correspondng average (mean from CDF) and coverage throughput (5-tle n CDF). Table 4 Probablty dstrbuton for number of scheduled users per RB under dfferent OThreshold value Number of scheduled users per RB 1 3 4 Threshold 0.1 0.46 0.683 0.071 0.001 0. 0.01 0.403 0.494 0.081 0.3 0.003 0.19 0.489 0.378 0.4 0.001 0.043 0.79 0.677 0.5 0 0.016 0.16 0.857 The results are generated under 8 antenna elements at base staton and SVD precodng. It can be found both average and coverage throughput are mproved wth threshold ncreased up to 0.4, snce each user can be allocated more RBs; however, coverage comes down f threshold s further ncreased to 0.5 as IUI becomes the unfavorable factor for cell edge users. Therefore, 0.4 seems to be a proper value and s thus assumed n the followng smulatons. C. Comparson wth sngle-user MIMO and MU- In Fgure 4 we compare the throughput performance of sngle-user MIMO and MU-LA. Ran adaptaton s enabled for sngle-user MIMO, so one user can have two data streams f better trouped can be provded than sngle stream transmsson. From the results t can be seen that mult-user MIMO wth MU-LA can lead to more than 30% throughput gan for both average and coverage, wth deal channel nowledge. However, n realstc case where ssrs and nsrs as descrbed n Secton II.A s consdered, the average gan can be stll as hgh as 6%, but the coverage gan decreases a lot to 7%. Fgure 4 Throughput comparson between SU-MIMO and MU-LA Next MU-LA s compared wth MU-, so that t can be observed under whch condton there could be gans n throughput from schedulng more than two users. Fgure 5 shows the comparson among sngle-user MIMO, MU- and MU-LA wth 4 antenna elements at base staton and wth both SVD and PMI based precodng. It should be noted that both MU- and MU-LA are schedulng two users at each RB, so the dfference only s how to select the secondary user. From Fgure 5 t can be seen that the throughput gan over sngle-user MIMO s ~10% for PMI based precodng and ~0% for SVD based precodng. MU- and MU-LA are gvng very smlar performance. Fgure 3 Throughput performance under dfferent OThreshold value Fgure 5 Performance comparson among SU, MU- and MU-LA wth 4 antenna elements at base staton 790
Fgure 6 shows smlar comparson as Fgure 5 but wth 8 transmt antennas. The gan of mult-user MIMO over sngleuser MIMO s larger. MU-LA provdes better average but slghtly worse coverage throughput than MU-. In partcular, f channel state nformaton can be accurately avalable at base staton (SVD based precodng s possble), t s meanngful to schedule more than two users per RB. The reason s that precodng wth 8 antenna elements can provde good separaton among scheduled users, so the mult-user dversty gan can be obtaned wth less sgnal penalty from ZF. Fgure 6 Performance comparson among SU, MU- and MU-LA wth 8 antenna elements at base staton D. MU-LA n mult-cell scenaro In ths subsecton, we wll evaluate the performance of MU-LA extended n a mult-cell scenaro. Intra-ste CoMP s consdered, and the modelng s lsted n Table 5 aganst snglecell mult-user MIMO. Table 5 Smulaton modellng of ntra-ste CoMP Sngle-cell MU- Intra-ste CoMP MIMO Number of base 3 1 statons per ste Number of users per 10 30 base staton Number of antennas 4/8 1/4 per base staton Maxmum coscheduled users 4 1 Path loss b/w base staton and UE 1 path loss value for all antennas 3 path loss values, each value for a sectorzed antenna group The throughput gan of ntra-ste CoMP over sngle-cell mult-user MIMO s shown n Table 6 for both 3GPP Case 1 and ITU UM channel models. The results show that the CoMP gan (by usng extended MU-LA) s not so attractve for 3GPP Case 1 channel, and the reason s that ICI from co-located base statons s already qute lmted n sngle-cell scenaro, so splttng some power to serve users n other co-located cells (wth hgh path loss) s not cost-effcent. On the other hand, n IUT UM channel the azmuth spread s much larger, and ICI s lmtng the system performance snce some porton of transmt power wll be leaed to co-located cells even wth perfect precodng. Intra-ste CoMP can effectvely mtgate such ICI. Table 6 Throughput gan of ntra-ste CoMP over sngle-cell MIMO 3GPP Case 1 UM 4-Tx 8-Tx 4-Tx 8-Tx Average gan (%) 8.4 13.0 50.7 57.4 Coverage gan (%) 0.6 8.5 30.0 V. CONCLUSIONS In ths paper, we evaluate the throughput performance of a specfc mult-user schedulng algorthm through system level smulaton based on LTE-A downln. The algorthm greedly adds secondary users accordng to the orthogonalty wth already scheduled users and the PF metrc of canddate users, and arbtrary number of users can be scheduled on a same RB. The scalablty and complexty s desrable from base staton mplementaton pont of vew. From smulatons under sngle-cell scenaro, the optmal value for orthogonalty threshold s found. Although the throughput performance s qute mpressve (~40% gan over sngle-user MIMO) wth 8 transmt antennas and perfect channel nowledge, the gan loos less attractve for 4 transmt antennas or mperfect precodng. From smulatons under mult-cell scenaro, the throughput gan of ntra-ste CoMP over sngle-cell mult-user MIMO s qute lmted for 3GPP Case 1 channel, whle qute sgnfcant for ITU UM channel. The layer adaptve schedulng algorthm can be used as a reasonable scheduler wth proper parameter adjustment. REFERENCES [1] Gesbert D., Kountours M., Heath Jr. R.W., Chae C.B., Sälzer T., Shftng the MIMO Paradgm: From sngle-user to multuser communcatons, IEEE Sgnal Processng Magazne, vol. 4, no. 5, Sept. 007, pp. 36-46. [] Care G., Shama S. (Shtz), On the achevable throughput of a multantennagaussan broadcast channel, Informaton Theory, IEEE Transactons on, vol. 49, no. 7, July 003, pp. 1691-1706. [3] Dmc G., Sdropoulos N.D., On downln beamformng wth greedy user selecton: performance analyss and a smple new algorthm Sgnal Processng, IEEE Transactons on, vol. 53, no. 10, Oct. 005, pp. 3857-3868. [4] Xun Luo, Yang Yang, Zhangcheng Dng, Shu Fang, Shaoqan L, Lower bound capacty based schedulng n MUMIMO system, n Proc. of Communcatons and Networng n Chna (ChnaCom), 011 6th Internatonal ICST Conference on, Aug. 011, pp. 191-195. [5] Batsta, R.L., dos Santos, R.B., Macel, T.F., Fretas, W.C., Cavalcant, F.R.P., Performance Evaluaton for Resource Allocaton Algorthms n CoMP Systems, n Proc. of IEEE 7th Vehcular Technology Conference (VTC), Sept. 010, pp. 1-5. [6] 3GPP TR 36.819, Coordnated mult-pont operaton for LTE physcal layer aspects (Release 11), Verson 11.1.0, Dec. 011. [7] Huang, H., Trvellato, M., Hottnen, A., Shaf, M., Smth, P., Valenzuela, R., Increasng downln cellular throughput wth lmted networ MIMO coordnaton, Wreless Communcatons, IEEE Transactons on, vol. 8, no. 6, June 009, pp. 983-989. [8] Pennanen, H., Haataja, T., Lenonen, J., Toll, A., Latva-aho, M., System level evaluaton of TDD based LTE-Advanced MIMO- OFDMA systems, n Proc. of GLOBECOM Worshops (GC Wshps), 010 IEEE, Dec. 010, pp. 809-813. [9] 3GPP R1-111301, Results for CoMP Phase I evaluaton n homogeneous networ, CMCC, Barcelona, Span, May 011. [10] Poharyal, A., et.al., HARQ Aware Frequency Doman Pacet Scheduler wth Dfferent Degrees of Farness for the UTRAN Long Term Evoluton, n Proc. of IEEE 65th Vehcular Technology Conference (VTC), Aprl 007, pp. 761-765. 791