Combned Beamformng and Schedulng for Hgh Speed Downlnk Packet Access Alexander Seeger Informaton and Communcaton Moble Semens A.G. Hofmannstr.5, 8359 Munch, Germany Alexander.Seeger@semens.com Marcn Skora Department of Electrcal Engneerng Unversty of Notre Dame Notre Dame, IN 46556, USA mskora@eee.org Wolfgang Utschck Inst. for Crcut Theory and Sgnal Proc. Munch Unversty of Technology 89 Munch, Germany utschck@nws.e.tum.de Abstract Jont operaton of Hgh Speed Downlnk Packet Access (HSDPA) and adaptve antennas n a WCDMA cellular network s consdered. The total throughput per cell of HSDPA depends heavly on the strategy employed by the scheduler. It s argued that the maxmum SIR scheduler, whch maxmzes throughput n the sngle antenna system by servng only the user wth best momentary channel qualty, wll not substantally beneft from adaptve antennas due to excess self nterference at the recever. As an alternatve a set of schedulng strateges s proposed, whch rely on servng smultaneously multple spatally separated users. System level smulatons show that wth the Spatal Dvson Multple Access scheduler t s possble to acheve almost twofold throughput mprovement over maxmum SIR when usng four antennas per sector. I. INTRODUCTION TO HSDPA IN CONJUNCTION WITH DOWNLINK BEAMFORMING Rapd varatons n the qualty of a rado channel (fadng) pose a problem for the dgtal transmsson of voce n a cellular network, due to the strct requrements on data rate and transmsson delay. One of the technques to deal wth changng channel, used e.g. n UMTS, s fast power control. The same approach can also be used for the transmsson of packet data; however, relaxed requrements on rate and delay of the latter make dfferent lnk adaptaton technques much more effcent. HSDPA, a part of UMTS standard responsble for downlnk transmsson of packet data, facltates Adaptve Modulaton and Codng (AMC), Hybrd Automatc Repeat Request (H-ARQ) and transmsson schedulng to acheve hgh data rates up to.8 Mbt/s per cell. The scheduler, a base staton entty responsble for dynamcal assgnment of rado resources to moble statons awatng data, uses ts knowledge about the momentary channel qualty to decde how many spreadng codes wll be assgned to each actve user. The maxmum SIR strategy,.e., assgnng n each transmsson tme nterval (TTI) all HSDPA spreadng codes to the user wth best momentary channel condton, guarantees hghest throughput per cell n a sngle antenna system. An equvalent strategy can be defned for systems wth multple antennas: n each TTI, the user wth the best channel qualty s determned, and served exclusvely wth the weght vector that maxmzes the beamformng gan. However, ths strategy has a consderable drawback. Snce the base staton drects almost the whole transmt power to a sngle user, the beamformng ncreases both receved sgnal power Ths work was supported n part by NASA grant NAG5-53. and self nterference. Effectvely, ths strategy does not sgnfcantly help users wth good channel qualtes, whch generally experence much hgher ntracell nterference than ntercell nterference. Snce these users are most often selected by the maxmum SIR scheduler, the performance ncrease due to applcaton of adaptve antennas s expected to be small. An addtonal dsadvantage of maxmum SIR scheduler used n conjuncton wth closely spaced adaptve antennas (.e., a grd of beams) s the generaton of rapdly changng spatal nterference patterns. Such effect s caused by the fact, that n each TTI most of the base staton power wll be radated n a dfferent drecton. Ths wll cause huge tme varatons of both ntra- and ntercell nterference, whch can serously mpact channel qualty estmaton at the moble statons. It s nterestng to note that the problem of ncreased ntracell nterference s not present, when adaptve antennas are used to smultaneously serve multple users wth dfferent weght vectors and fractons of the total base staton power. However, servng multple users smultaneously reduces the selecton gan [5],[6]. Wthn ths paper, schedulng strateges offerng a compromse between both objectves are proposed. The content s structured as follows. System and propagaton models are dscussed n Secton. Secton 3 descrbes the concept of jont schedulng and beamformng and proposes two scheduler algorthms, wth ther performance evaluated n Secton 4 based on smulaton results. Fnally, Secton 5 draws the conclusons. II. SYSTEM MODEL A. Propagaton model The large-scale model of a cellular system consdered n ths paper assumes a hexagonal grd of cells wth unform user dstrbuton and base statons wth three sectors per ste. To avod boundary effects, two rngs of adjacent base statons are smulated but only the data from the central cell s evaluated. Sectorzaton s acheved by drectonal antennas. The average channel attenuaton from each BS to each MS follows exponental path loss and lognormal shadowng. Based on the average receved power (proportonal to the drectonal antenna gan and nversely proportonal to the channel attenuaton) each MS performs a cell selecton emulatng a standard handover process. GLOBECOM 3-5 - -783-7974-8/3/$7. 3 IEEE
Each sector s served wth a unform lnear array of N ant closely spaced antennas. The channel between the BS and MS s subject to frequency selectve fadng, modeled n baseband as tap delay lne wth N tap temporal taps wth coeffcents followng zero-mean complex Gaussan dstrbuton. Every channel coeffcent vector h m from all BS antenna elements for the m-th temporal tap s spatally correlated and H R h m = E h h equals [ ] [ ] m m π [ m] = L L a ( φ φ ) a H avg m ( φ φ ) fφ ( φ ) R h dφ, () π where L avg s the average path attenuaton (due to path loss and shadowng), L m s the relatve power of the m-th tap and a(φ), φ and f Φ (φ) are the antenna array steerng vector, drecton of departure and the power azmuth spectrum, respectvely. An mportant smplfcaton s the assumpton of nterference-lmted system. Such assumpton can be made for urban scenaros, where the cells are small, base statons are placed densely, and ntercell nterference always domnates over thermal nose at the moble staton recevers. B. Transmsson wth multple antennas There are three approaches usng multple antennas at the BS for downlnk communcaton standardzed n WCDMA: uplnk-based beamformng (grd of beams), closed-loop transmt dversty and open-loop transmt dversty (space-tme codng). It s nterestng to note that dversty technques do not offer a consderable beneft for HSDPA, snce ther purpose s reducton of channel qualty varatons due to fadng, whch s the source of schedulng gan. Thus only the grd of beams approach s consdered here. The transmtter employs CDMA to send a superposton of M parallel data streams. Baseband sgnal s (d) (t) conveyng d-th data stream, carres β (d) fracton of total transmt power P T, and uses spreadng factor SF (d) and antenna weght vector w (d). The total transmtted baseband sgnal s M () t = PT w ( d ) β ( d ) s( d )() t x. () d = The ntercell nterference, just as the usable sgnal, s affected by path loss, shadowng, antenna gan and fadng to each of the nterferng stes. In such case, the radaton pattern of k-th nterferng base staton sector can be descrbed by a transmsson covarance matrx R x (k), H [ ] ( k ) ( k ) () ( k E x t x ) () t R x =, (3) where x (k) (t) s the sgnal transmtted from all antenna elements of Node-B and sector k. The overall receved ntercell nterference power can be calculated as a sum of receved power from all non-servng Node-B sectors: ( k ) N N = sectors taps I nter k= = ( k ) H ( k ) ( k ) h R x h. (4) C. RAKE recever and SIR calculaton Each moble staton s assumed to use an dealzed RAKE recever wth perfect channel knowledge, one RAKE fnger per channel temporal tap, and deal maxmum rato combnng (MRC). Wth these assumptons t s possble to calculate the SIR n the de-spread complex data sequence leavng MRC unt. The complex coeffcent between the d-th data stream and -th RAKE fnger can be calculated as ( d ) H f = w ( d ) h PT β ( d ). (5) If the j-th RAKE fnger s usng the spreadng code of the data sgnal δ, then all data sgnals receved at other temporal tap than j are source of ntracell nterference, ( δ ) SF ( ) ( ) f δ δ j SIR j = N taps N d ( d ) f + I = d = j nter After maxmum rato combnng the overall SIR s N taps. (6) ( δ ) ( δ ) SIR MRC = SIR j, (7) whch can be used to approxmate the WCDMA modulator, beamformer, channel and RAKE recever as an equvalent AWGN channel (on a TTI-wse bass). D. Lnk-level performance of HSDPA The dependency between SIR and throughput has been establshed by lnk-level smulatons n []. For the assumed subset of allowed modulaton and codng schemes (QPSK wth rates ¼, ½, ¾ and 6-QAM wth rates ½, ¾), the lnk level performance s shown n Fg.. From these throughput curves, the convex hull s taken to determne the acheved data rate as a functon of sgnal-to-nterference rato. Note that these lnk level results already nclude possble retransmssons. The SIR value s assumed to be constant for subsequent hybrd ARQ retransmssons. III. j= COMBINED SCHEDULING AND BEAMFORMING A. Spatally separatng scheduler The SIR at the recever lstenng to one of the spreadng codes consderably depends on the weght vectors appled to ths code (nfluence on usable sgnal power), as well as vectors appled to the remanng codes (nfluence on ntracell nterference). The followng strategy, subsequently called spatally separatng scheduler, attempts to serve several spatally separated users smultaneously n each TTI. In the grd of beams scenaro, the scheduler consders two users to be spatally separated f they are served wth non-overlappng beams. The number of separated users can never exceed the number of antennas. In the frst step, all moble statons are parttoned nto groups wth the same servng antenna weght GLOBECOM 3-5 - -783-7974-8/3/$7. 3 IEEE
Throughput per spreadng code (n kbt/s) 8 7 6 5 4 3 6-QAM, R=3/4 6-QAM, R=/ QPSK, R=3/4 QPSK, R=/ QPSK, R=/4-5 - -5 5 5 Sgnal-to-Interference Rato per spreadng code (n db) Fgure. Lnk level performance of selected modulaton and codng schemes. vector or beam. Wthn each group, the user wth best channel qualty s determned (unless the group s empty), and the remanng users are not further consdered for transmsson. As a next step, the scheduler constructs all possble sets of spatally separated users (sets wth only one user are also consdered), and tres to determne, whch of them acheves the hghest total throughput. After the best set s determned, the 5 HS-DSCH spreadng codes are splt among the users contaned n ths set. Determnng the best set of users s complcated by the fact, that the scheduler does not exactly know how the assgnment of transmt power among beams wll affect the nterference levels of the users. We wll follow an approach n whch scheduler neglects ths problem and assumes that the SIR at every Rake recever s lnearly proportonal to the sgnal power (.e., assumes that sgnal power does not affect ntracell nterference). Thus, for each group of K users wth sgnal-to nterference ratos SIR (k), k =,, K sgnaled over CQI feedback, the scheduler calculates the followng metrc: J = K k = ( ) log + SIR( k )/ K. (8) Ths metrc corresponds to the sum of Shannon capactes to the selected users under the smplfyng assumpton mentoned above. B. Space Dvson Multple Access scheduler The scheduler presented n the prevous secton serves multple spatally separated users smultaneously. Although the condton used to conclude spatal separaton s qute smple, t promses a reducton n ntracell nterference especally n the grd of beams scenaro. Ideally, spatal separaton leads to complete elmnaton of such nterference. In such a case, t s not necessary to use only a subset of 5 HS-DSCH spreadng codes, but a full set of codes could be reused for each moble staton. Of course, n order to keep the total transmt power unchanged, the amount of power per spreadng code must be lowered. However, performng a communcaton on a larger bandwdth (wth spreadng codes correspondng to bandwdth) s more power-effcent n terms of Shannon capacty, and the code reuse can brng consderable throughput mprovement. The scheduler mplementng ths strategy performs a Spatal Dvson Multple Access (SDMA). The selecton process of the set of users to be served s based on the same prncple as the spatally separatng scheduler. The only dfference comes n the way of calculatng the metrc for the set of K users: J = K k= ( ) log + SIR( k )/ K. (9) Snce n practce the spatal separaton s not perfect, reuse of spreadng codes can lead to cross-talk between two data channel usng the same spreadng and scramblng code. Ths mght be undesrable, snce f both channels use the same channel code, the recever mght have a problem n dstngushng both channels. Ths problem can be mtgated by usng multple scramblng codes n a sngle cell, and servng each user wth dfferent scramblng code and the full set of spreadng codes. Both varants of SDMA scheduler (wth sngle or multple scramblng codes) wll be consdered for the grd of beams system. IV. SIMULATION RESULTS The basc system smulaton approach s a snapshot smulator. The complete set of smulaton parameters s shown n Table. Performance of schedulng strateges consdered for adaptve antenna systems s evaluated for two man scenaros: grd of 8 beams for 4 antennas and grd of 4 beams for antennas. Fg. and Fg. 3 present smulaton results for the grd of beams scenaro wth four antennas and eght beams, at both Pedestran A and Vehcular A delay spread profles. The TABLE I. Parameter SYSTEM SIMULATION PARAMETERS. Value Number of sectors per ste 3 Drectonal antenna pattern see UMTS 3.3 Adaptve antenna scenaros Antenna spacng.5 wavelength Path loss exponent 3,76 Shadowng standard devaton 8 db Ste-to-ste shadowng correlaton 5% Sector-to-sector shadowng correlaton % User placement Unform Servng ste selecton Mnmum path loss + shadowng Power delay profle Vehcular A, Pedestran A Recever RAKE wth one fnger for each tap, perfect MRC Fracton of Node-B power dedcated to 8% HS-DSCH Number of HS-DSCH spreadng codes 5 Number of snapshots Angular spread for drecton of departure Laplace dstrbuted wth deg standard devaton GLOBECOM 3-5 - -783-7974-8/3/$7. 3 IEEE
Average throughput n kbts/s.5 x 4.5.5 Grd of 8 beams, 4 antennas, Pedestran A, 3 sectors per ste Sngle antenna - Maxmum SIR Grd of beams - Maxmum SIR Grd of beams - Spatal separaton Grd of beams - SDMA multple scramb. Grd of beams - SDMA sngle scramb. Monte Carlo PDF of SIR.9.8.7.6.5.4.3.. Pedestran A, 3 sectors per ste Sngle antenna Fgure. Average throughput per cell for dfferent scheduler types for Pedestran A power delay profle. - -8-5 - -9-6 -3 3 6 9 5 8 SIR n db Fgure 4. Dstrbuton of SIR for Pedestran A power delay profle. Average throughput n kbts/s 8 6 4 8 6 4 Grd of 8 beams, 4 antennas, Vehcular A, 3 sectors per ste Sngle antenna - Maxmum SIR Grd of beams - Maxmum SIR Grd of beams - Spatal separaton Grd of beams - SDMA multple scramb. Grd of beams - SDMA sngle scramb. Monte Carlo PDF of SIR..8.6.4...8.6.4 Vehcular A, 3 sectors per ste Sngle antenna. Fgure 3. Average throughput per cell for dfferent scheduler types for Vehcular A power delay profle. average throughput per cell for all consdered schedulers s shown n comparson to a sngle antenna scenaro wth the maxmum SIR scheduler. For both power delay profles, the applcaton of 4 antennas allows the maxmum SIR scheduler to ncrease the throughput by a constant offset (except for MCS saturaton range), compared to sngle antenna case. Ths s due to the beamformng gan whch ncreases the power of the usable sgnal by about 6 db (4 antennas), but t also ncreased the ntracell nterference. Snce ntracell nterference s more predomnant n Vehcular A, ts SIR and throughput ncrease s lower than n Pedestran A. Ths effect s better llustrated n the SIR dstrbutons before schedulng for sngle antenna and 4 antennas as presented n Fg. 4 and Fg. 5. In upper range of SIR the dstrbuton shfted up by 4 db n Pedestran A and only by - -8-5 - -9-6 -3 3 6 9 Sgnal-to-nterference rato n db Fgure 5. Dstrbuton of SIR for Vehcular A power delay profle. db n Vehcular A (whch s a very small beneft from havng 4 antennas). In Pedestran A, the SIR ncrease quckly caused throughput saturaton at the maxmum value allowed wth hghest MCS, ndcatng a need for even hgher order modulatons. The spatally separatng scheduler turns out to be not consderably better than maxmum SIR scheduler. In comparson between both, spatally separatng scheduler gans by reducng the ntracell nterference, but loses some of the multuser gan. The balance of the two effects s postve for the Vehcular A, but negatve for Pedestran A. Compared to the prevous strateges, the SDMA schedulng offers a tremendous performance gan. Usng the full set of spreadng codes per user almost doubles the average throughput n Vehcular A. In Pedestran A the gan s even hgher, snce the reducton of SIR (caused by splttng the same GLOBECOM 3-53 - -783-7974-8/3/$7. 3 IEEE
Average number of users served smultaneously 4 3.5 3.5.5.5 Sngle antenna Fgure 6. Number of users served smultaneously by SDMA scheduler for one, two, and four antennas, and the power delay profle Pedestran A. transmsson power over larger number of codes) elmnates the MCS saturaton pont. The throughput wth SDMA schedulng ncreases much faster wth the number of users than wth maxmum SIR. Responsble for ths s a new type of multuser gan, called here spatal multuser gan. Ths gan comes from ablty of SDMA scheduler to fnd larger and larger sets of mutually spatally separated users, whch can be then served smultaneously. Ths s llustrated n Fg. 6, whch shows the average number of smultaneously served users versus total number of users per cell. Fnally, Fg. and Fg. 3 reveal that between two SDMA strateges, the varant wth multple scramblng codes was consequently worse than the one wth sngle scramblng code. The reason for ths s an excess nterference comng from the common channels (usng the remanng % of the Node-B transmt power), whch are dstrbuted evenly across all eght beams. Wth one scramblng code, these channels pose mnmal nterference to HS-DSCH, but wth multple they nterfere wth HS-DSCH spreadng codes located on overlappng beams usng dfferent scramblng codes. V. CONCLUSIONS The man conclusons from the smulaton results presented are as follows. Usng maxmum SIR scheduler wth adaptve antennas ncreases the power of the usable sgnal, but also ncreases the ntracell nterference. The latter effect s partcularly harmful n scenaros wth wde power delay profle n system wth 4 antennas only db gan n SIR was observed. In scenaros wth narrow power delay profle SIR reached hgh values, whch could not be fully exploted due to lmted range of modulaton and codng schemes. Adaptve antennas can be used to separate users spatally. The spatally separatng scheduler uses ths ablty to counteract the ncrease of ntracell nterference experenced by maxmum SIR. It acheves ths by sharng the HS-DSCH codes between multple spatally separated users, at a cost of lower multuser gan (selectng multple nstead of sngle best user). The net balance s postve only for wde power delay profle scenaros. The true beneft of spatal separaton s enabled by the SDMA scheduler. By reusng all spreadng codes for spatally separated users Spatal Dvson Multple Access takes advantage of addtonal bandwdth provded by spatal dmenson. As an ndrect, but consderable beneft, SDMA scheduler rases the maxmum throughput permtted by the hghest MCS by the factor equal to number of antennas. Both effects allowed the SDMA scheduler to almost double the average throughput per cell, compared to maxmum SIR. In a scenaro wth and Pedestran A, the average throughput reaches up to 5 Mbt/s for 5 users per cell. REFERENCES [] 3GPP, TR, v3.., Selecton procedures for the choce of rado transmsson technologes of the UMTS, 998. [] M. Döttlng, J. Mchel, B. Raaf, Hybrd ARQ and adaptve modulaton and codng schemes for hgh speed downlnk packet access, IEEE Internatonal Symposum on Personal, Indoor and Moble Rado Communcatons, pp. 73-77,. [3] K.I. Pedersen, P.E. Mogensen, B.H. Fleury, A Stochastc Model of the Temporal and Azmuthal Dsperson Seen at the Base Staton n Outdoor Propagaton Envronments, IEEE Trans. Vehcular Technology, vol. 49, no., pp. 437-447, Mar.. [4] 3GPP, TR 5.877 Techncal Report on Beamformng Enhancements. [5] D.J. Mazzarese and W.A. Krzymeń, Hgh throughput downlnk cellular packet data access wth multple antennas and multuser dversty, n Proc. IEEE Vehcular Technology Conference, pp. 79-83, Aprl 3. [6] H. Vswanathan and K. Kumaran, "Rate schedulng n multple antenna downlnk wreless systems", n Proc. Allerton Conference on Communcatons and Control, October. GLOBECOM 3-54 - -783-7974-8/3/$7. 3 IEEE