Exponential Effective SIR Metric for LTE Downlink

Transcription

1 Exponental Effectve SIR Metrc for LTE Downlnk Joan Olmos, Albert Serra, Slva Ruz, Maro García-Lozano, Davd Gonzalez Sgnal Theory and Communcatons Department Unverstat Poltècnca de Catalunya (UPC) Barcelona, Span emal: [juan.jose.olmos, slva.ruz, maro.garca-lozano, Abstract 3GPP LTE s the evoluton of UMTS whch wll make possble to delver next generaton hgh qualty multmeda servces accordng to the users expectatons. The flexblty of the downlnk OFDM rado nterface wth Adaptve Modulaton and Codng (AMC), MIMO and H-ARQ plays a crucal role n achevng the low latency and hgh spectral effcency promsed by the new rado access standard. Ths paper presents a LTE DL lnk level smulator whose man purpose s to generate sutable look-up tables to nterface wth a system level smulator. In ths context, the Exponental Effectve SIR (EESIR) metrc s a lnk abstracton model that s used to properly characterze multstate channels. The reference BLER curves n AWGN channel and the parameters of the EESIR model are gven for the complete lst of CQI s specfed for LTE DL. The obtaned results also nclude curves of mean lnk level throughput for dfferent AMC formats and MIMO confguratons. 3GPP Long Term Evoluton; Lnk level smulator; Exponental Effectve SIR I. INTRODUCTION 3GPP LTE s the evoluton of UMTS whch wll make possble to delver next generaton hgh qualty multmeda servces accordng to the users expectatons []. OFDM/OFDMA have been selected by 3GPP as the physcal layer and multple access schemes for DL LTE. OFDM can delver a hgh bt rate even n a heavy multpath envronment thanks to the use of many narrow-band subcarrers whch are easy to equalze. Furthermore, snce fadng s almost flat wthn each subcarrer, sutable MIMO schemes can easly be appled on a subcarrer bass. OFDM shows an nherent frequency dversty effect f a coded block s sent on a set of subcarrers spannng a bandwdth hgher that the channel coherence bandwdth. Adaptve modulaton and codng (AMC) allows delverng a reasonable throughput to the users at the cell border whle achevng hgh spectral effcency for users near the enb. Another key aspect of the LTE DL rado nterface s the use of fast retransmsson at the MAC level wth ncremental redundancy (H-ARQ). H-ARQ smoothes the AMC throughput curves, thus allowng less frequent swtchng between AMC formats. In ths paper we descrbe an adhoc LTE DL lnk level smulator, entrely wrtten n C++, whch can generate the sutable look-up tables for the EESIR model to nterface wth a LTE system level smulator. A methodology to tran the EESIR model usng the lnk level results s explaned and the reference BLER curves n AWGN channel and the parameters of the EESIR model are gven for the complete lst of CQI s specfed for LTE DL. Results showng mean lnk level throughput vs. mean SIR are gven as well. The paper s organzed as follows: secton II s a descrpton of the developed smulator, secton III presents the results focusng on average performance fgures, secton IV descrbes the methodology for tranng the EESIR model and dscusses the results and secton V ncludes the conclusons and future work. II. DESCRIPTION OF THE E-UTRA DL LINK LEVEL SIMULATOR In order to feed a system level smulator wth the lnk level performance, an ad-hoc lnk level smulator has been programmed n C++ language. The E-UTRA DL lnk level smulator features an OFDM physcal layer n accordance wth [2] and [3], (see Fg. ). The rate /3 turbo encoder wth varable code block sze creates three ndependent streams wth systematc and redundant bts. Those streams are nterleaved and fed to the crcular buffer rate matchng and H-ARQ procedure. Varable codng rate s acheved by applyng dfferent puncturng patterns dependng on the current H-ARQ redundancy verson (RV). Up to four ncremental redundancy (IR) transmssons per code block are allowed. The turbo decoder uses a MAP algorthm and a maxmum of 8 decodng teratons. ACK/NACK error free transmsson s consdered. The QAM modulator generates the complex modulated symbols, belongng to ether a QPSK, 6QAM or 64QAM constellaton. Sgnallng and plot symbols overhead s not consdered. After the IFFT and the cyclc prefx (CP) addton, the tme-doman samples are processed by the stochastc MIMO multpath moble channel smulator. In SISO mode the moble channel smulator can smulate any desred power delay profle (p.d.p.) (ncludng Rce dstrbuted LOS propagaton path). The Doppler spectra of the dfferent paths are shaped by a classcal Jakes low-pass flter. In MIMO mode the channel smulator uses a stochastc matrx channel model that ncludes the antenna correlaton effects. All elements of the matrx channel are ndependently generated as for the SISO channel accordng to the same p.d.p.. Correlaton among antennas s ntroduced usng the procedures recommended n [4]. Based on ths assumpton, the spatal correlaton matrx of the MIMO rado channel s the Kronecker product of the spatal correlaton matrx at the recever and transmtter. Ref. [4] also specfes three dfferent levels of antenna correlaton, termed low, medum and hgh correlaton level. The antenna Ths work s supported by Spansh Natonal Scence Councl under grant TEC C02-02

2 correlaton parameters from [4] have been appled to obtan the results presented n ths paper. Ideal channel estmaton s assumed for channel equalsaton (one tap equalzer n SISO mode) and for MIMO processng at the recever. The smulator ncludes MIMO spatal multplexng modes wth lnear recevers (Zero Forcng and MMSE) as well as Alamout space frequency codng transmsson dversty. In order to speed up smulatons, the lnk level smulator s splt n two ndependent programs. The frst stage ncludes the wdeband OFDM modulaton/demodulaton and the multpath channel. The output of ths stage s the set of the log-lkelhood ratos (LLR) of the receved encoded bts. The second stage flters the set of LLRs, whch account for all the smulated OFDM subcarrers, to extract the LLRs of the Physcal Resource Blocks (we consder PRB = 2 subcarrers x ms) allocated to the smulated user. In ths way the tme and frequency correlatons of the channel response, as seen by the moble termnal, are properly captured. The set of LLRs are used n the second smulaton stage to feed the turbo decoder wth the soft relablty values of the receved bts. Input from LA BS Random bt generator Rate /3 turbo encoder MS BLER statstcs and lnk-level throughput calculaton Turbo decoder corresponds to a typcal Raylegh envronment BER curve. MMSE recever slghtly outperforms ZF recever. The multpath channel model consdered n Fg. 2 s Extended Pedestran A (EPA), as specfed n [5], wth a 3km/h pedestran speed. TABLE I. LINK LEVEL SIMULATOR PARAMETERS Parameter Value Carrer frequency 2 GHz Smulated Bandwdth 20 MHz Sub-carrer spacng 5 khz OFDM PHY parameters Cyclc Prefx of 4.69 μs FFT sze 2048 Number of useful sub-carrers 200 OFDM symbol duraton 7.43 μs Number of sub-carrers per 2 PRB Number of allocated PRBs TTI nterval ms Number of useful OFDM symbols per TTI AWGN, EPA channel model at Power Delay Profle Pedestran speed (3 km/h) and Extended Typcal Urban (ETU) channel model [5] Channel Codng Turbo code basc rate /3 Rate Matchng and H-ARQ Accordng to [3]. Max 4 IR transmssons. AMC formats for tranng EESIR model As specfed n [8] for each of the 5 possble CQI ndexes Channel estmaton Ideal Antenna scheme SISO and MIMO L H-ARQ & rate matchng L H-ARQ & de-rate matchng.e+00.e-0 QAM constellaton mappng QAM constellaton de-mappng.e-02 MUX and MIMO Tx processng MUX and MIMO Rx processng Uncoded BER.E-03 IFFT & CP addton Multpath MIMO channel CP removal & FFT Fg. Block dagram of one MIMO layer of the E-UTRA DL lnk level smulator III. LTE DL LINK AVERAGE THROUGHPUT The set of parameters used n the smulatons are shown n TABLE I. In order to test the valdty of the smulator results, several curves of the uncoded BER were frst obtaned. Fg. 2 shows the mean uncoded BER vs. mean SIR, for QPSK and 6QAM wth MIMO 2x2 n spatal multplexng mode and n Alamout transmt/receve dversty mode. When usng the Alamout scheme, two antennas at the moble node are used to mplement an order 2 MRC recever dversty scheme. Ths mode gves a BER slope correspondng to order 4 dversty. Ths s n contrast wth the slope for an Alamout 2x scheme (also shown n Fg. 2) that corresponds to order 2 dversty. Also n Fg. 2 the BER slope for the MIMO spatal multplexng modes.e-04.e-05.e-06 QPSK MIMO 2X2 ZF QPSK MIMO 2X2 MMSE 6QAM MIMO 2X2 Alamout TD/MRC QPSK MIMO 2x2 Alamout TD/MRC QPSK MIMO 2x Alamout TD SINR (db) Fg. 2. Uncoded BER vs. mean SIR for dfferent modulatons and MIMO confguratons n EPA multpath 3km/h In order to obtan results on mean lnk level achevable throughput, the bandwdth allocated to the smulated user has been set to PRB and the smulated turbo code block szes are the smaller ones specfed for E-UTRA DL (nevertheless the smulator tme resoluton corresponds to a 20MHz bandwdth). The obtaned lnk level throughput can thus be consdered the E-UTRA DL baselne performance, snce hgher code block szes and the nherent frequency dversty of usng a set of PRB s wll lead to hgher throughput fgures. The mean throughput s evaluated as the product of the modulaton spectral effcency (n bts/s/hz) tmes the code rate and the H-ARQ retransmsson effcency. The SIR s averaged over fast fadng but not over

3 shadowng fadng, whch s properly modelled at system level smulaton. Fg. 3 and Fg. 4 show the E-UTRA DL throughput, for dfferent modulatons and code rates, wth 2x2 MIMO spatal multplexng mode (MMSE recever) and EPA multpath channel at a pedestran speed of 3km/h. Each of these fgures assumes a dfferent correlaton level between antennas. It can be verfed that the antenna correlaton has an mportant mpact on the achevable throughput wth MIMO. Fg. 5 compares the mean throughput for Alamout Tx/Rx dversty vs. MIMO spatal multplexng wth MMSE processng. For low correlaton MIMO channel, spatal multplexng gets hgher throughput than Alamout/MRC f average SIR s hgher than 0 db. For medum and hgh correlaton MIMO channel the throughput obtaned by Alamout/MRC, below an average SIR of 2 db and 27 db respectvely, s hgher than the obtaned by MIMO/MMSE. So, as t s usually recognzed, spatal multplexng mode makes more sense at hgh SIR. addtonal penalty, n the form of a nose ncrease, can be ntroduced to take nto account the resdual channel dstorton due to mperfect channel estmaton/equalzaton and the possble transmtter dstorton (error vector magntude). The only obstacle s that, n a wdeband multpath channel envronment, each OFDM subcarrer may suffer a dfferent attenuaton. The transmtted code blocks are thus affected by a multstate channel. In the lterature, [6] [7] [9], the EESIR lnk abstracton model s ntroduced to account for multstate channels whle stll beng able to use the AWGN BLER vs. SIR curve as a reference for the BLER estmaton under real condtons. Lnk Level Troughput [bt/s/hz] QPSK - /3 QPSK - /2 QPSK - 2/3 QPSK - 4/5 6QAM - /2 6QAM - 2/3 6QAM - 4/5 64QAM -2/3 64QAM - 4/ QPSK - /3 Lnk Level Troughput [bt/s/hz] QPSK - /2 QPSK - 2/3 QPSK - 4/5 6QAM - /2 6QAM - 2/3 6QAM - 4/5 64QAM - 2/3 64QAM - 4/5 0 Fg. 4 E-UTRA DL 2x2 MIMO MMSE lnk level throughput wth H-ARQ n multpath EPA 3km/h and hgh antenna correlaton 0,00 MIMO 2x2 MMSE (Low Correlaton) MIMO 2x2 MMSE (Medum Correlaton) MIMO 2x2 MMSE (Hgh Correlaton) 2 8,00 MIMO 2x2 Alamout/MRC (Low Correlaton) MIMO 2x2 Alamout/MRC (Medum Correlaton) 0 Fg. 3 E-UTRA DL 2x2 MIMO MMSE lnk level throughput wth H-ARQ n multpath EPA channel 3km/h and medum antenna correlaton Lnk Level Troughput [bt/s/hz] 6,00 4,00 MIMO 2x2 Alamout/MRC (Hgh Correlaton) IV. TRAINING THE EESIR MODEL FOR MULTISTATE CHANNELS The classcal approach for nterfacng lnk level to system level smulators s to generate look-up tables of mean coded BER versus the mean SIR values. Ths approach s no longer vald for wdeband packet mode systems lke LTE, where one transport block, made up of one or more code blocks, s send n a TTI of ms usng a varable number of physcal resource blocks. In LTE each physcal resource block occupes 2 OFDM subcarrers (80 khz). Due to the short duraton of one TTI, wth the excepton of hgh speed vehcular envronments, the wdeband channel transfer functon wll undergo only lttle varatons durng the transmsson of a sngle code block. Under ths condtons, the BLER vs SIR curve for a gven transport format (a gven combnaton of modulaton, code block sze and code rate) could be approxmated by the BLER curve for that transport format combnaton under AWGN channel condtons. An 2,00 0,00 Fg. 5 E-UTRA DL Alamout vs. spatal multplexng lnk level throughput wth H-ARQ n multpath EPA 3km/h For a gven multstate channel wth N dfferent SIR measurements {γ,γ 2,,γ N }, the EESIR s defned as the value, γ ef, that accomplshes the equaton, [9]: γ ef N γ k = ln e () N k = where N depends on the frequency resoluton of the measurements avalable at the moble node. At the moble

4 node, knowledge of the EESIR metrc for each resource block group (RBG) s requred n order to report the CQIs specfed by LTE, so n practce the value of N should span a bandwdth equvalent to one RBG. A RBG s a set of, 2, 3 or 4 consecutve resource blocks (the actual value depends on the system bandwdth, see [8]). The parameter allows adjustng the expermental BLER measurements to the reference AWGN BLER curve. Once s known the reference curve can be used to obtan an estmaton of the BLER under any multstate channel for whch a set of measurements s avalable. By usng eq. (), the EESIR concept summarzes all these measurements n a sngle scalar value. The value of, specfc for a gven modulaton, mult-antenna/dversty confguraton, code block sze and code rate, must be obtaned by tranng the EESIR model wth results from the lnk level smulator. The nformaton passed from the lnk level to the system level smulator s a set of look up tables, specfc for each transport format, where each look-up table ncludes the correspondng value of plus the specfc AWGN reference BLER curve. Ths curve must be obtaned by smulatng the complete transmsson chan, ncludng codng and nterleavng, under AWGN assumptons. Based on the CQI ndexes gven by the LTE standard, [8], only 5 dfferent transport formats need to be consdered for the downlnk. The system level smulator wll then use these tables to compute the set of CQI values to be reported by each moble node to the enb for each RBG. The enb needs the CQI s to perform the schedulng of downlnk transmssons based on the qualty reported by each moble node n each RBG. Notce that, accordng to [8], the reported CQI ndex for a gven group of resource blocks, should allow the recepton of a sngle transport block wth a BLER not hgher than 0. after the transmsson of the frst H-ARQ RV. Ths means that the relevant BLER curves must be obtaned assumng the specfc code puncturng pattern that s used for the frst H-ARQ RV. Fg. 6. Several realzatons (snapshots) of the ETU multpath channel frequency response showng dfferent frequency selectve fadng dynamc range. Wdeband mean SNR=dB. Subcarrer spacng=5 khz Next we descrbe the procedure that has been followed to obtan the value of for each transport format combnaton. The frst stage s to obtan the reference BLER curves under AWGN channel (sold lnes n Fg. 7). Notce that the reference BLER curves are almost regularly spaced n steps of 2dB EESIR ncrease. The second stage s to generate a bg number of snapshots of the moble multpath channel wth known set of measurements: {γ,γ 2,,γ N}, where γ k (k=,..,n) denotes the SIR for the subcarrer k and the snapshot. For each nstance of the multpath channel a bg number of code blocks are smulated and a BLER measurement for the snapshot s obtaned. We call BLER to ths measurement. The channel transfer functon remans constant durng the smulaton of a gven channel snapshot. Snce eq. () doesn t capture the frequency correlaton of selectve fadng, a random sub-carrer nterleaver s used so that frequency selectve fadng does not affect always the same bts n the code block. To obtan, the EESIR of the snapshot s expressed as a functon of the unknown,.e.: γ ef = ln N N γ k e (2) k = then, f we call M to the number of smulated channel snapshots, the value of s obtaned as: M = arg mn ( ) = ( BLER ) 2 BLERR γef (3) where BLER R (γ) s the reference BLER curve for AWGN channel. That s, to obtan we mnmze the mean squared error between the estmated BLER and the measured BLER for the set of smulated snapshots usng numercal methods. Fg. 7 shows the fttng of EESIR model for E-UTRA DL BLER (RV=0) wth SISO confguraton and RBG sze=. The BLER measurements (scattered markers n Fg. 7) have been obtaned usng the ETU channel, [5], where the delay of the last propagaton path has been taken equal to 4.69μs (nstead of 5μs) to ensure that t falls wthn the cyclc prefx nterval. Ths channel has been selected because t shows hgh frequency selectvty (see Fg. 6) but the resultng values of are ndependent of the p.d.p. used for tranng the model. TABLE II lsts the code block szes, the obtaned and the resdual r.m.s. error for each CQI. The code block szes are the maxmum that allow fttng the entre encoded block n a sngle TTI. For CQI 2, and gven the small code rate, one TTI has not enough capacty to transmt a sngle code block of the lowest sze. In these cases we have assumed that the transport block can use addtonal subcarrers n the neghborng PRB. TABLE II apples only for RBG=, snce for RBG> the allowable code block szes would be hgher and the EESIR model would need retranng. It can be notced n Fg. 7 and TABLE II that the expermental to estmated BLER fttng degrades for the CQIs related to 64QAM. The dervaton of the EESIR concept, whch can be found n [6], s based on approxmatng the BLER as a functon of the sgnal to nose ratos at the code symbol level, whle the EESIR defnton

5 appled n ths paper (see eq. ()) reles on the subcarrer SIRs (at modulaton symbol level). Snce 64QAM entals unequal protecton aganst nose for the encoded bts, the statstcs of the LLRs at the code symbol level, depend on the subcarrer SIR but also on the consdered modulaton symbol and on the weght of the encoded bt wthn that symbol. So the subcarrer SIR alone can t properly predct the BLER and the EESIR concept shows hgher r.m.s. error. When the code rate s close to unty, the decson reles bascally on the ntrnsc nformaton (voltage of the decson varables) and BLER -(-BER) n, where n s the encoded block sze and BER s approxmately /6 of the modulaton symbol error probablty, whch s a functon of the SIR and s qute unform over the set of modulaton symbols (snce SIR 9dB). Ths explans the relatvely good ft of EESIR model for CQI 5. TABLE II. VALUES OF AND R.M.S. ERROR FOR THE COMPLETE SET OF CQI S ( SISO CONFIGURATION AND RBG SIZE=) CQI Modulaton Code rate Turbo code block sze r.m.s. error QPSK QPSK QPSK QPSK QPSK QPSK QAM QAM QAM QAM QAM QAM QAM QAM QAM Fg. 7. E-UTRA DL BLER (RV=0) vs EESIR for SISO confguraton (RBG sze=) V. CONCLUSIONS AND FUTURE WORK A lnk level smulator for LTE DL has been descrbed. The smulator allows obtanng the sutable look-up tables to nterface a system level smulator by means of the EESIR model. The mean throughput curves for dfferent MIMO and AMC combnatons have also been dscussed. In future work, the EESIR reference BLER curves and parameters wll be computed also for RBG of sze 2, 3 and 4 and for MIMO confguratons. ACKNOWLEDGMENT Ths work s supported by Spansh Natonal Scence Councl under grant TEC C REFERENCES [] 3GPP TR 25.84, Physcal Layer Aspects for E-UTRA (Release 7), v7..0 [2] 3GPP TS 36.2, E-UTRA Physcal Channels and Modulaton (Release 8), v8.5.0 [3] 3GPP TS 36.22, E-UTRA Multplexng and Channel Codng (Release 8), v8.5.0 [4] 3GPP TS 36.0, E-UTRA UE Rado Transmsson and Recepton, (Release 8), v8.4.0 [5] 3GPP TS 36.04, E-UTRA Base Staton (BS) rado transmsson and recepton (Release 8), v8.4.0 [6] 3GPP2-C , Effectve-SNR Mappng for Modelng Frame Error Rates n Multple-state Channels, Ercsson [7] K.Bruennghaus, D.Astdlyt, T.Slzert, S.Vsur, A.Alexou, S.Karger, G.Seraj, Lnk Performance Models for System Level Smulatons of Broadband Rado Access Systems, PIMRC 2005 [8] 3GPP TS 36.23, E-UTRA Physcal layer procedures, (Release 8), v8.5.0 [9] 3GPP TR , Feasblty Study for Orthogonal Frequency Dvson Multplexng (OFDM) for UTRAN enhancement, (Release 6), v6.0.0