Adaptive TDMA/OFDMA for Wide-Area Coverage and Vehicular Velocities

Transcription

1 Adaptive TDMA/FDMA for Wide-Area Coverage and Vehiular Veloities Mikael Sternad *, Sorour Falahati *, Tommy Svensson, and Daniel Aronsson * * Signals and Systems, Uppsala University, P Box 8,SE-71 0 Uppsala, Sweden, Dept. Signals and Systems, Chalmers University of Tehnology, SE-196 Göteborg, Sweden, {mikael.sternad, sorour.falahati, daniel.aronsson}@signal.uu.se; tommy.svensson@s.halmers.se Abstrat Within the EU FP6 Integrated Projet WINNER, adaptive transmission is investigated as a key tehnology for boosting the spetral effiieny of a new radio interfae for G systems. Adaptive alloation of time-frequeny hunks in an FDM-based system offers a signifiant potential, but also poses hallenges. Within work pakage two of WINNER, we study ritial issues suh as the feasibility of adaptive transmission over fading downlink/uplink hannels to/from vehiular terminals, the orresponding required hannel predition auray, and the required feedbak ontrol bandwidth. This paper summarizes reent results obtained within WINNER, and related results obtained within the Swedish Wireless IP projet. Index Terms G mobile wireless systems, adaptive transmission and multiple aess, rthogonal Frequeny Division Multiplexing (FDM), spetral effiieny. A I. INTRDUCTIN DAPTIVE systems alloate (shedule) time, frequeny and antenna resoures based on hannel quality and user requirements. They enable effiient resoure utilization and multi-user sheduling gains, when hannels to different terminals fade independently. In systems based on time division multiple aess/ adaptive FDM (TDMA/FDMA), time-frequeny resoures (hunks) are alloated. This provides a flexible small-sale granularity of the resoures, ideal for transmitting small as well as large pakets. Based on the results obtained within the Swedish Wireless IP projet 1, we are assessing the feasibility of suh methods in novel broadband radio interfaes within the EU FP6 Integrated Projet WINNER. We here investigate adaptive downlinks and uplinks based on fast sheduling and link adaptation, also for users at vehiular speeds, with a non-adaptive fall-bak mode for very fast moving users. The non-adaptive fallbak mode design is outside the sope of this paper, but should be based on time, frequeny and spae diversity tehniques. Alloation of fast 1 This work has been performed in the framework of the IST projet IST WINNER, whih is partly funded by the European Union. The authors would like to aknowledge the ontributions of their olleagues. fading hannels requires hannel predition. The signal to interferene and noise ratio (SINR) is to be predited for all potential resoures in future transmission. In the proposed downlink, eah terminal predits the SINR over a major part of the total bandwidth. All ative terminals report soure oded SINR values or soure oded suggested modulation formats over a shared uplink ontrol hannel. A resoure sheduler, loated lose to one or several radio aess points, alloates the downlink resoures. In an adaptive uplink, one has the problem that hannels from eah potential user will have to be estimated and predited. In a system using frequeny division duplex (FDD), the estimation must be arried out in the aess point, and has to be based on pilots transmitted by all ative terminals. To avoid unaeptable pilot overhead, these pilots must be transmitted simultaneously, by overlapping pilots. This is related to the problem of estimating hannels from multiple antennas []. There are four key problems in the system, and we investigate solutions to the first two of them in this paper: 1. Prediting the short-term fading over a large fration of the wavelength at high arrier frequenies and high terminal veloities.. btaining a high hannel predition quality of uplinks for many ative users with a low pilot overhead.. Ahieving a low hannel state feedbak data rate in a system with large bandwidth and fine granularity of the hunks.. Maintaining a good frequeny synhronization of all uplinks to avoid signifiant interarrier interferene. II. FDD DWNLINK AND UPLINK DESIGN We explore adaptive TDMA/FDMA designed at arrier frequeny GHz. Adaptive transmission to vehiular users over TDMA/FDMA downlinks has earlier been investigated in [1] and [] for a more narrowband system of MHz bandwidth at arrier frequeny 1.9 GHz and early results in the WINNER ontext with a different design are given in []. The basi time-frequeny resoure unit is denoted hunk. It onsists of a retangular time-frequeny area that omprises a number of subsequent FDM symbols and a number of

2 adjaent subarriers, and is alloated exlusively to one user data flow. In Table I, we show the assumed hunk size along with important system parameters. A hunk ontains payload and pilot symbols. It may also ontain ontrol symbols to minimize feedbak delays, i.e. in-band ontrol signaling. The number of offered payload bits in a hunk depends on the utilized modulation and oding sheme (MCS), whih is seleted adaptively. TABLE I BASIC PARAMETERS FR FDD WIDE-AREA DWNLINK AND UPLINK Centre frequeny.0 +/- 0.8 GHz Number of FDM sub-arriers 10 FFT BW 0.0 MHz Signal BW 16. MHz paired Number of used subarriers 8 Sub-arrier spaing 191 Hz FDM symbol length (exl. CP) 1.0 µs Cyli prefix (CP) length.00 µs Physial hunk size 16. khz x 7. µs Chunk size in symbols 8 x 6 = 8 A. Downlink The downlink is designed as follows. Regular pilot patterns are transmitted on the downlink. Based on hannel measurements up to hunk time i, all ative terminals predit the hannel quality in all hunks within a sub-band of interest at the future hunk time i+. These reports are soure-oded and transmitted on uplink ontrol symbols within the uplink hunks at time i+1. The appropriate MCS that ould be used by eah terminal in eah hunk is then determined based on SINR rate limits. The adaptive resoure sheduler at the aess point alloates eah hunk at time i+ exlusively to one of the flows. The alloation is reported by ontrol symbols in the alloated downlink hunks at time i+. In eah downlink hunk, four pilot symbols and eight inband ontrol symbols are assumed. MCSs ranging from BPSK rate ½ to 6-QAM rate /6 are used. Thus, the number of payload bits per hunk may vary between 18 and 180. Let the subarriers within a hunk be enumerated from =1-8, and let the six FDM symbols be enumerated from s=1-6, see Fig. 1 (left). Channel preditors utilize subarriers and 7, and the hannel estimation an utilize the ontrol symbols in a deision-direted mode. Symbols (,s)=(,1),(,),(7,1),(7,) arry the downlink ontrol bits, and determines whih of the present hunks that belong to whih flow. Coded -QAM symbols are used. Symbols (,s) = (,) and (7,) are known unoded -QAM pilots. At multi-antenna aess points, different pilots are transmitted from eah antenna. The pilots are used for two purposes: hannel estimation for oherent detetion within hunks with payload of interest and predition for all frequenies of interest for future adaptive downlink transmission. When FDM symbol within hunk i has arrived, hannel predition for hunk i+ is performed. The required predition horizon to the end of hunk i+ is. x 0.7 ms = 0.8 ms. Freq =1 s=1 FDD Downlink: C C P C C P C C P C C P P: pilots C: Control Freq Time FDD Uplink: : verlapping. pilots Fig. 1. Pilot and ontrol symbol patterns in FDD downlink and uplink. : downlink ontrol feedbak (-8 symbols) Symbols (,s) = (,), (,), (7,) and (7,) arry ontrol information for the next uplink transmission telling whih of the next uplink hunks have been appointed to whih uplink flow. Coded -QAM symbols are used. Symbols (,s) = (,6) and (7,6) are known unoded -QAM pilots, used for oherent detetion and for updating the preditor states. Control information is broadast and has to be detetable by all users. Thus, their reeption range essentially defines the ell radius. The ontrol symbols are all loated on the same subarriers as the pilots, to make use of them for deisiondireted hannel estimation, in order to redue the hannel estimation mean square error (MSE) and the predition MSE. B. Uplink n the uplink, terminals taking part in adaptive transmission are in ompetition for a part of the total 16. MHz band, alled a ontention band. All ative terminals assigned to a ontention band simultaneously send overlapping pilot signals during hunk time i. All eight symbols within FDM symbol s= as shown in Fig.1 (right) are reserved for this purpose. Preditors loated at the aess point predit the hannels for all terminals at time i+. The predition is based on the latest and previously reeived signals at the loations of the overlapping pilots. The appropriate MCS that ould be used by eah terminal in eah hunk is then determined. The adaptive resoure sheduler assigns the uplink transmission for time i+ and informs the terminals by in-band signaling using ontrol symbols of the downlink hunk at time i+1. The in-band ontrol symbols on the uplink whih are positioned early in the hunk, see Fig. 1 (right), are part of the ontrol loop for the downlink. They arry the downlink hannel predition reports from all ative terminals. Their number (depending on the number of ative terminals and their veloities) an be adapted to the requirements, varying from 1 to 8 oded -QAM symbols per hunk. The use of overlapping pilots in a Kalman filter that simultaneously estimates and predits all hannels is desribed in [6]. In multi-antenna reeivers, the predition should be performed separately for all reeiving antennas within a setor/ell. The predition horizon from symbol in hunk i to the end of hunk i+ is. x 0.7 ms = 0.8 ms. Time

3 III. C HANNEL PREDICTIN The feedbak loops for the FDD system is designed to be as fast as possible, under realisti onstraints imposed by omputation times and signaling delays. However, hannel predition is needed for sheduling and link adaptation, sine extrapolating the present hannel estimate would lead to large performane losses. Extensive investigations of hannel power preditors were performed in e.g. [7], [8] and [9]. Both theoretial analyses and evaluations on a large set of measured hannels with MHz bandwidth were taken into aount. It was onluded that the lass of hannel power preditors that performed best on measured data was based on linear predition of the omplex baseband hannel, followed by use of a quadrati unbiased preditor to predit the hannel power. The noise level was found to be the ruial limiting fator for the attainable performane auray. Shemes that utilize many samples to average and suppress noise will provide better predition performane. It is therefore advantageous for the predition performane if a large fration of symbols within the subarriers are either pilots, or an be used for deisiondireted estimation, as disussed also in setion II. In [7], the most signifiant taps of the hannel impulse response are predited in the time domain. Here, we instead do the predition in the frequeny domain. A set of linear predition filters, eah responsible for its own subband of the total bandwidth, is utilized. The state spae algorithm desribed in [] is used to predit the omplex hannel and the unbiased quadrati preditor is used to predit the hannel power. The algorithm in [] starts by deriving a Kalman preditor. The preditor utilizes the orrelation of the hannel in the frequeny domain by prediting p pilot-ontaining subarriers in parallel. It also utilizes the orrelation in the time domain of the fading hannel. The number p is a ompromise between performane and omputational omplexity. We use p = 8, spanning hunks. This means that 6 suh Kalman estimators would be required to over a omplete band of 10 hunks in both ases. In [] it is shown that the present FDM hannel predition problem is ideally suited to the appliation of a novel lowomplexity approximation of the Kalman algorithm, the Generalized Constant Gain algorithm [10], whih avoids the need to update a quadrati state-spae Riati differene equation, responsible for the dominant omputational load in Kalman algorithms. Autoregressive models of order are used to model the hannel orrelation in time. They are adjusted to the fading statistis. The state update equations of the Kalman/GCG estimators are based on these models. In FDD uplinks, hannel predition is performed at the aess point, using overlapping pilots from the terminals. A generalization of the Kalman algorithm of [], desribed in [6], is used. Its performane is investigated below. Comparative evaluations of the frequeny domain and the time domain approah are urrently underway within the Swedish Wireless IP projet. Fig.. Normalized predition error, as a funtion of the predition horizon saled in arrier wavelengths, and as funtion of the SINR. Results for FDD downlink, full duplex terminals, over WINNER Urban Maro hannels, with a Kalman algorithm utilizing 8 subarriers. TABLE II WINNER URBAN MACR PWER DELAY PRFILE τ n = {0, 10, 0, 0, 60, 80, 60, 70, 8, 100, 10, 106, 70, 70, 760, 600, 610, 6} ns p n = {,., 6.98,.718, 6.98, 8.698,.0, 7.0, 9.00, , , , 1.016, 1.716, , 1.01, 17.71, 19.81} db IV. PERFRMANCE RESULTS A. Frequeny domain Kalman/GCG hannel predition The FDD downlink with WINNER Urban Maro power delay profile, defined in Table II, and white noise with known power are used in the investigation of the predition error as a funtion of the predition horizon, saled in wavelengths, for different values of the SINR. Fig. shows the results for full duplex terminals using all timeslots for updating the preditor with measurements. There is a large sensitivity to SINR, and the predition error grows with the predition horizon. The results for predition horizon zero represent the filter NMSE. Fig. shows results for FDD uplinks, where users and 8 users respetively are simultaneously transmitting overlapping pilots, all having the same average reeived power. In a Kalman estimator based on overlapping pilots, separate sets of states are used for desribing the hannel of eah user. The autoregressive models that desribe the fading statistis of eah user are adjusted individually to the veloity of the users [6]. Uplink ontrol information ould be used for improving the estimate by deision-direted methods, but it is not used in the presented results. All users have the same veloity and travel through the same type of propagation environment, but with independent hannel realizations. These results are based on the ITU Vehiular-A hannel model. The average reeived power is assumed equal for all users (slow power ontrol). The results indiate that predition based on overlapping pilots will derease in auray with an inreasing number of terminals, but this derease is rather modest. Channel preditions in FDD uplinks in whih not too many users oupy eah ontention band thus seems feasible.

4 the predition horizon D = 0.8 ms as required in setion II for both downlink and uplink. It is evident that adaptive transmission an be expeted to work in the widest variety of situations in the wide-area FDD downlinks, whereas adaptive transmission in the wide-area FDD uplink, an work in many important situations. TABLE III ESTIMATES F THE SINR LIMITS FR ADAPTIVE TRANSMISSIN AT GHZ WITH PREDICTIN HRIZN IN WAVELENGTHS SINR, predition horizon 0 km/h 0 km/h 70 km/h Downlink < 0 db, 6 db, 1. db, Uplink, users 0 db, 7 db, 1 db, Uplink, 8 users. db, 11 db, 0 db, Fig.. Normalized predition error as a funtion of the predition horizon saled in arrier wavelengths, and as funtion of the SINR. Results for FDD uplink over ITU Vehiular-A hannels, with a Kalman algorithm for overlapping uplink pilots utilizing 8 subarriers. Result for (top) and 8 (bottom) simultaneous users per ontention band. B. Limits for adaptive FDD TDMA/FDMA transmission The predition auray depends on the predition horizon h saled in wavelength, whih in turn depends on the veloity v, the predition horizon in time D and the arrier wavelength λ via the relation h = vd / λ. The predition auray also depends on the SINR. Thus, adaptive transmission to/from a terminal will be feasible up to a maximal veloity for a given SINR, or equivalently, down to a limiting SINR at a given veloity. Estimates of the limiting SINR values are given here, based on the design in setion II and the results in Fig. -. They are onservative, sine the predition is performed to the far end of the hunk to be alloated. The predition auray to less distant symbol loations will be higher. From earlier investigations of the sensitivity for MCS rate limits to predition errors, it has been found that if the rate limits are adjusted to take the predition unertainty into aount, a predition NMSE of 0.1 for an unoded system leads to only a minor degradation in the spetral effiieny [11], [1], but for oded shemes the sensitivity to predition errors is slightly larger. We here use an upper limit of 0.1 for the allowed normalized variane of the omplex predition error. Table III shows the resulting limits for the SINR along with the orresponding predition horizons in wavelengths for V. SIMULATIN RESULTS FR FDD DWNLINK In this setion, the SINR limits stated in Table III are ompared to the simulation results of the adaptive FDD TDMA/FDMA downlink using the WINNER Urban Maro model. In the multilink simulations, all hannels have the same statistial properties, all terminals are full duplex and have the same veloity. The interferene is modeled with white Gaussian noise and all terminals have the same average SINR. The sheduling strategy used is Proportional Fair, whih in this ase, where all users have the same average SINR, redues to the Max Throughput strategy of giving the hunk to the user who an use the highest modulation-oding rate. The resoure sheduling buffers are never emptied. The hannels are not perfetly flat within the hunks: there is in general variability both in the time diretion and in the frequeny diretion. Within eah hunk, the modulation and oding sheme potentially used by eah user is determined by taking the average predited SINR, SINR av, and the predited SINR at the worst point within the hunk, SINR w, for that user. The weighted average is used as the effetive SINR: SINR[ db] = bsinr av [ db] + ( 1 b) SINRw[dB]. The parameter b an be used to tune the performane of the sheme when we have signifiant hannel variability within the hunks. With b=1, large variability leads to a large inrease in the BER, sine the properties of the worst orner of the hunk generates most errors. With b=0, we obtain a onservative sheme, that tends to provide on average less errors than the target BER. In all results shown, b=0. is used. The effet of hannel estimation errors on the demodulation is not onsidered. As sheduling unit (SU), we use 1-bit pakets (small IP paket). The overhead due to CRC ode and sequene numbers for link ARQ is not taken into aount. Eah SU is distributed among the alloated hunks, and superfluous payload symbols are filled with zeros. Separate MCS is used for eah hunk. If all bits belonging to a SU are reeived orretly, it is released to higher layers. therwise, a link

5 retransmission would our. However, link level retransmission is not used in the simulations. Coded M-QAM are used with eight rates: BPSK rate ½, QPSK rate ½, QPSK rate ¾, 16-QAM rate ½, 16-QAM rate /, 16-QAM rate /6, 6-QAM rate / and 6-QAM rate /6, based on the rate ½ onstraint length 9 onvolutional ode with generator polynomials (61,7) in otal representation, whih is puntured to obtain the higher rate odes. The rate limits are optimized under a maximal bit error rate onstraint of 0.001, for a given average SINR and predition error variane. The atual average bit error rate beomes lower, sine the maximal BER is targeted for the MCS limits. The throughput is defined as the number of payload bits in orretly reeived SUs divided by the total number of transmitted payload symbols. The effet of different terminal veloities and the orresponding predition unertainties on the throughput, the multiuser sheduling gain and the bit error rate are measured. Table III indiates that when all users have either 10 db SINR or 19 db SINR, the adaptation sheme should work rather well for all veloities up to 70 km/h at 19 db, but diffiulties may be enountered at 70 km/h when the SINR is 10 db. Fig. onfirms this statement. The dashed urves show the performane in the presene of predition inauray. The solid urves show the ase when preditions are assumed perfet, but the effet of the hannel variability within hunks due to the fading in time is taken into aount. For the dashed urves, the MCSs are designed to attain the BER onstraints in the presene of predition errors. This goal is fulfilled, with one exeption: that of 70 km/h at 19 db SINR. The orresponding paket error rates for the 1 bit SU is below 1%, whih indiates that performane ould be improved by tuning the sheme more aggressively. With an inreasing speed, and orrespondingly inreasing predition unertainty, the rate limits are tuned more onservatively, and the throughput is dereased. Signifiant multiuser sheduling gains are however preserved also with rather long preditions. At 70 km/h and 10 db SINR the sheme fails due to too high predition unertainty, and this operating point is beyond the limit given in Table III. Note also that the predition unertainty dereases with the number of ative users sine with many users, hunks are given to users with relatively good hannels having a small predition unertainty. With the results above, we have shown that it is possible to adaptively utilize the short-term fading also for vehiular users at suh high arrier frequenies as GHz. REFERENCES [1] W. Wang, T. ttosson, M.Sternad, A. Ahlén and A. Svensson, Impat of multiuser diversity and hannel variability on adaptive FDM, VTC 00 Fall, rlando, FL, t. 00. [] M. Sternad, T. ttosson, A. Ahlén and A. Svensson, Attaining both overage and high spetral effiieny with adaptive FDMA downlinks, VTC 00-Fall, rlando, Fla, t. 00. [] G. Auer, Analysis of pilot-symbol aided hannel estimation for FDM systems with multiple transmit antennas, IEEE ICC 0, Paris, June 00. Fig.. Throughput as a funtion of the number of ative users, all with the same average SINR of 10 db (top) and 19 db (bottom) in FDD wide-area downlink. Solid urves take hannel variability within hunks into aount but neglet the predition unertainty. Dashed urves take predition unertainty into aount. [] M. Sternad and D. Aronsson, Channel estimation and predition for adaptive FDM downlinks," IEEE VTC 00-Fall, rlando, Fla, t. 00. [] M. Sternad, S. Falahati and T. Svensson. Adaptive FDMA/TDMA Transmission at Vehiular Veloities. Pro. Wireless World Researh Forum WWRF1, Toronto, Canada, Nov 00. [6] M. Sternad and D. Aronsson, Channel estimation and predition for adaptive FDMA/TDMA uplinks, based on overlapping pilots, International Conferene on Aoustis, Speeh and Signal Proessing (ICASSP 00). Philadelphia, PA, USA, Marh nline: [7] T. Ekman, Predition of Mobile Radio Channels. Modelling and Design. Ph.D. Th., Signals and Syst., Uppsala Univ. [8] M. Sternad, T. Ekman and A. Ahlén, Power predition on broadband hannels, IEEE Vehiular Tehnology Conferene VTC01-Spring, Rhodes, Greee, May [9] T. Ekman, M. Sternad and A. Ahlen, "Unbiased power predition on broadband hannels" IEEE VTC 00-Fall, Vanouver, Canada, Sept. 00. [10] M. Sternad, L. Lindbom and A. Ahlén, "Wiener design of adaptation algorithms with time-invariant gains," IEEE Transations on Signal Proessing, vol. 0, pp , August 00. [11] S. Falahati, A. Svensson, T. Ekman and M. Sternad, "Adaptive modulation systems for predited wireless hannels," IEEE Trans. on Communiations, vol., Feb. 00, pp [1] S. Falahati, A. Svensson, M. Sternad and H. Mei, "Adaptive Trellisoded modulation over predited flat fading hannels," IEEE VTC 00- Fall, rlando, Fla, t. 00.