On the Definition of Reference Scenarios for LTE-A Link Level Simulations within COST IC1004

Transcription

1 EUROPEAN COOPERATION IN THE FIELD OF SCIENTIFIC AND TECHNICAL RESEARCH IC1004 TD(13)06043 Málaga, Spain 6-8 February, 013 EURO-COST SOURCE: UPC - Universitat Politècnica de Catalunya (Spain) On the Definition of Reference Scenarios for LTE-A Link Level Simulations within COST IC1004 Joan Olmos, Albert Serra and Silvia Ruiz Escola d Enginyeria de Telecomunicació i Aeroespacial de Castelldefels (EETAC-UPC) Esteve Terradas, Castelldefels, SPAIN Phone: Fax: olmos@tsc.upc.edu

2 On the Definition of Reference Scenarios for LTE-A Link Level Simulations within COST IC1004 Joan Olmos, Albert Serra and Silvia Ruiz Universitat Politècnica de Catalunya, WiComTec group, Abstract The simulation of LTE and LTE-A at link level is a research topic for several COST IC1004 institutions. It was agreed in previous meetings that it would be useful to define a common set of scenarios for LTE and LTE-A link level simulations with the goal to allow researchers to configure the different simulation tools in a way that the obtained results can be properly compared and simulators become calibrated to a common reference. Researchers working at system level would also benefit from this initiative by having access to well established link level results and look-up tables. This paper addresses the task by identifying a set of topics, features and parameters to consider when programming the link level simulators. The proposed aspects derive from the experience gained while building our own link level simulator at UPC, but since the LTE radio interface is extremely flexible and supports many features, it is difficult to be exhaustive. The aim is to focus on the relevant topics for link level simulator calibration and to discuss the trade-off between simulation complexity and compliance with the standards. The detailed definition of the link level scenarios is left for future joint work, in collaboration with the interested IC1004 institutions. Keywords LTE, Link level simulator, Link to system mapping 1. Introduction The simulation of mobile communication systems is usually divided in two different instances, i.e., link and system level simulators. While the link level simulator simulates a single radio link with full details, the task is too complex to be extended to simulate the whole system. The system level simulator takes into account a complete cell deployment and relies on simplified look-up tables generated offline by the link level. The border between link and system level, i.e., which tasks are performed at each simulator, depends on the characteristics of the standard and on the simulator designs. For example, the modelling of multipath frequency selective fading used to be a link level task, but with the LTE scheduling resolution of 15 khz (in bandwidth) and 1 ms (in time), multipath fading must be also considered at the system level. The modelling of ARQ is another task that used to be restricted to link level and now (with Hybrid- ARQ) has merged with the system level. These examples show that the interface between link and system level simulators is becoming increasingly complex. The introduction of OFDMA has created the need for link abstraction techniques to predict the Transport Block (TB) error rate (BLER) under frequency selective fading conditions, while the short TTI of 1 ms, which is usually shorter than the channel coherence time, makes short term fading variations almost irrelevant, thus allowing the modelling of the turbo code error performance through AWGN curves. The fact that the LTE-A PHY layer can be configured in many different ways (called transmission modes ) also adds complexity to the simulators. In this paper we focus on the tasks to perform at the LTE-A link level simulator and discuss the trade-off between simulation complexity and compliance with the standards. The aim is to start the collaboration among the interested institutions and eventually reach a consensus on which link level scenarios are worth to analyze and compare results in the frame of the COST action. Once that consensus is reached, a comprehensive database, with relevant link to system mapping parameters for the different transmission modes, could be created and used as a reference for researchers working at system level. 1

3 . Scope and objectives of the link level simulator Considering that the link level simulator is a complex project by itself, the first logical step would be to identify which are the objectives and specifications of the project. A few high level decisions on the scope of the desired results are useful to set research priorities. Among those bullet points (which are not mutually exclusive) we have: 1. Simulator configuration: 1.1. LTE (Releases 8-9), LTE-A (Releases 10-) 1.. FDD, TDD modes 1.3. Downlink (DL), Uplink (UL) 1.4. Unicast, Multicast (e-mbms) 1.5. Set of supported LTE / LTE-A transmission modes 1.6. Simulated bandwidth (contiguous or distributed RBs. Frequency hopping?) 1.7. MIMO matrix channel simulation: purely stochastic, geometry-based stochastic, measurement campaign traces 1.8. Number of antennas. Scope of the desired results:.1. Uncoded BER.. BLER (with coding) in AWGN conditions for SISO mode and single HARQ transmission for the different AMC schemes (reference BLER curves).3. BLER and throughput (for relevant transmission modes) under realistic channel conditions (indoor, pedestrian, high speed, etc.) and for different redundancy versions of HARQ (BLER after receiving 1,,3 or 4 incremental redundancy versions) for the different AMC schemes.4. Training of link abstraction (LA) metrics: EESM, MIESM (LA includes HARQ?).5. Guidelines for transmission mode selection and AMC scheme selection 3. Simulator implementation: 3.1. Programming language: C++, Matlab, ns-, etc. 3.. Open access or closed. Open source or proprietary code Since the number of aspects to consider is quite high, a prioritization of the most interesting configurations and results is needed. Also a sequential implementation of the simulator features, from low to higher complexity, helps to reduce the debugging time by building new features over already tested software blocks. In this paper we deal with the FDD mode, since this is the background of our simulator. 3. Simulating the E-UTRA user plane Fig. 1 shows the mapping from transport channels to physical channels in E-UTRA, [1]. Usually only the user plane (PDSCH and PUSCH) is simulated at the link level simulator, but the control plane also must be taken into account, since a fraction of the PHY layer capacity is reserved for the control channels and this has an impact on the AMC schemes that can be applied for a given allocated bandwidth. Fig. 1. LTE mapping from transport to PHY channels

4 Another aspect to consider is the role of the pilot symbols (called reference signals in LTE). Pilot symbols must be considered, since they also reduce the available PHY capacity for the user plane (mainly for MIMO modes), and can also be simulated in detail if we need to test realistic channel estimation methods. As the 3GPP specifications recognise, there are two main processing chains to consider at link level: a) The transport channel processing for DL/UL-SCH, [1] b) The physical channel processing for PDSCH/PUSCH, [] Transport Block Size Determination Transport Block Generation and CRC attachment Code Block Segmentation and CRC attachment Turbo Coding Rate Matching and subblock interleaving Circular Buffer Code Block Concatenation Codeword to Layer Mapping Fig.. Transport channel processing (transmitter side) in LTE link level simulator where the physical channel is the immediate lower layer for the transport channel. Fig. depicts the generic block diagram of the transport channel processing at the transmitter side of the link level simulator. The receiving side does the reverse functions in the reverse order. The processing for UL or DL user plane is very similar. Fig. 3 depicts the physical channel processing for DL and SU-MIMO. Scrambling Scrambling Modulation Mapper Modulation Mapper Layer Mapper MIMO Precoding Resource Element Mapper Resource Element Mapper OFDM Signal Generation OFDM Signal Generation MIMO matrix channel model AWGN noise OFDM Demodulation OFDM Demodulation Resource Element De-mapper Resource Element De-mapper MIMO Receiver Processing SM->ZF, MMSE; Open or closed loop; SIC (if enabled) TD + MRC Modulation De-mapper Modulation De-mapper Layer Demapper Descrambling Descrambling Bit Level LLR Computation Bit Level LLR Computation MIMO Channel and SNR estimator MIMO post processing SINR per subcarrier Compute RI, CQI and PMI Fig. 3. Physical channel processing in LTE link level simulator (DL and SU-MIMO) 3

5 The processing for UL is very similar, the main difference being the additional DFT/IDFT blocks that are needed to map/de-map the transmitted signal (SC-FDMA) to the RB assigned to the UE. In the next sections we discuss in more detail the simulation of those two processing chains. 4. Simulating the transport channel processing The transport channel processing deals with transport block segmentation (to fit the maximum systematic block size of the turbo code: 6144 bits), turbo encoding/decoding, rate matching (to adapt from the rate 1/3 of the mother code to the desired rate specified by the applied AMC scheme) and interleaving, management of the HARQ processing (4 redundancy versions) and mapping to the MIMO layers. In LTE a codeword is a transport block after being processed by the chain of Fig.. A maximum of two codewords can be transmitted at once using MIMO spatial multiplexing modes. If the number of available MIMO layers is greater than two then parallel transmission is applied to increase the throughput. The most complex block in this processing chain lies in the turbo decoding at the receiver side. The transport block size must be determined according to the procedures of [3]. In Table 1 we show the DL transport block size for an allocated bandwidth of 5 RBs and SISO mode (1 layer/codeword). There are 6 resource elements reserved for pilots in each subframe (1 ms). In Table 1 the first column is the I MCS index, which is signalled in a 5 bit field within the PDCCH. From that field the UE extracts the modulation order and the TB size (by using tables indexed by the I TBS index). The PDSCH capacity (in bits/subframe) is then: ( ) C = Q N N 1 (14 N ) N [bits/subframe] (1) PDSCH m RB L PDCCH RS where Q m =,4,6 [bits/symbol] is the modulation order, N RB is the number of allocated RBs, N L is the number of layers available per codeword, N PDCCH is the number of OFDM symbols used for PDCCH and N RS is the number of RE reserved for pilots (per RB) within a subframe. If there are several spatial multiplexing (SM) MIMO layers available for the same codeword the PDSCH capacity is increased accordingly. If transmission diversity is applied then we should take N L =1 in expression (1). In order to compute the effective code rate, which is the key parameter for the rate matching stage, the next step is to compute the size of the payload (systematic bits) to fit inside the PDSCH capacity (already computed). The PDSCH payload, in systematic bits, is equal to the size of the TB (as derived from table , [3]) plus the size of the additional CRC fields. Due to restrictions in the internal interleaving, the size of the systematic turbo code block cannot be arbitrary; in fact there is a table of valid code block sizes ranging from 40 to 6144 bits, [1]. If the TB size is not bigger than 610 bits then code block fragmentation is not needed. In this case a single CRC of size 8, 16 or 4 bits is added to the TB. If the TB size is greater than 610 bits then a 4 bit CRC field is appended to the TB and the resulting packet is fragmented into several code blocks of (almost) equal length. Each one of the code block fragments is appended with an additional CRC of 4 bits. This allows the receiver to stop decoding the remaining fragments (if one of them fails to decode) and ask for the transmission of a new HARQ redundancy version (rv). In Table 1 column B is the PDSCH payload size, column C is the number of fragments and column K+ is the size of the turbo code block. The effective code rate (in bold in Table 1) is equal to the PDSCH payload over the PDSCH capacity. Notice that for I MCS 0 to 4 the effective code rate is less than the mother code rate (1/3). In this case all the coded bits are transmitted (no puncturing at all) and the effective code rate is reduced by bit repetition, that is, some of the coded bits are transmitted more than once within the same rv, thus leading to an energy gain when the receiver combines the LLRs belonging to the same coded bit. 4

6 Table 1. Transport block size for 5 RB, 1 layer/codeword and PDSCH OFDM symbols per subframe PDSCH Eff. BICM Mod. Pilots Capacity TB Size CRC Code Threshold Threshold IMCS Order ITBS [RE] [bits] [bits] [bits] B B' C C+ K+ C- K- F rate [db] delta [db] , -7, ,15-6,47 1, ,18-5,57 0, ,3-4,31 1, ,9-3,07 1, ,36-1,93 1, ,4-1,01 0, ,50 0,1 1, ,56 0,91 0, ,64 1,93 1, ,3,06 0, ,35,64 0, ,40 3,49 0, ,46 4,53 1, ,5 5,53 1, ,58 6,51 0, ,6 7,16 0, ,41 7,59 0, ,43 7,89 0, ,49 9,14 1, ,53 9,97 0, ,57 10,76 0, ,61,55 0, ,67 1,75 1, ,7 13,76 1, ,75 14,38 0, ,81 15,64 1, ,84 16,31 0,67 In case that the PDSCH payload cannot be exactly split into several code blocks of valid size, some filling bits are added at the beginning of the first code block in the PDSCH payload. Those filling bits are used at the turbo encoder to generate the redundancy but are not transmitted inside the systematic block. In fact they are known a priori by the receiver (they are all at logical zero) so the decoder assumes ideal LLR values for those bits. The available PHY layer room left by the filling bits is used to transmit additional redundant bits, thus reducing the effective code rate and increasing the probability of correct decoding, see Fig. 4. This situation arises, for example, when we use I MCS =0 with 1 RB. In this case the PDSCH payload consists of only one code block of size 40 bits and, assuming a CRC of 16 bits, the room left for the TB is 40-16=4 bits. Fig. 4. Processing of the filling bits But since, according to table , [3], the size of the TB is only 16 bits, we need to add 8 filling bits at the beginning of the systematic (information) bits. The PDSCH capacity is C PDSCH = 1 1 (1 (14-3)-6)=5 bits and the effective code rate is 3/5=0.17, since, from the point of view of the decoder the PDSCH payload size is 3 bits (16 from the TB and 16 from the CRC). With the notation of Fig. 4, for this example we have F=8, N=3 and M=60. 5

7 Given the effective code rate and the modulation order it is possible to compute the BICM SNR threshold for a particular AMC format, see Table 1. The BICM SNR threshold is the minimum SNR that would be required, with that AMC format, to achieve error free transmission with a capacity achieving error correcting code. The BICM threshold is computed by mapping the effective code rate through the BICM capacity curve of the modulation, see [4]. Fig. 5. LTE turbo code BLER (AWGN channel) vs. code block size In practice, the BLER curve for the LTE turbo codes (in AWGN channel) begins to fall down for a SNR slightly higher than the BICM threshold. So knowing that threshold is useful to check the validity of the obtained BLER curves in AWGN channel and to have a rough estimation of the minimum SNR needed for the particular AMC format. The offset of the BLER curve with respect to the BICM threshold, as well as the slope of the curve, depend on the size of the turbo code block. Fig. 5 shows the BLER curves for the full range of valid turbo code block sizes in AWGN channel with BPSK modulation and rate 1/3. As shown in Table 1, increasing the I MCS by one leads to a BICM threshold approximately 1 db higher than the previous format, see also Fig. 6 and Fig. 7. Fig. 6. BLER curves in AWGN channel for all I MCS (5 RBs) Fig. 8 shows the effects of code block fragmentation on the BLER curves for AWGN channel. It can be verified that the BLER with fragmentation fits quite well with a model of independent errors per code block. In Fig. 8 ncb is the number of code blocks (number of fragments) and P e means the BLER for ncb=1. 6

8 Fig. 7. BLER curves in AWGN channel for all I MCS (1 RB) Fig. 9 summarizes the AWGN performances, from an information transfer point of view, of all the I MCS formats in LTE for 1 RB. The circle (QPSK), triangle (16QAM) and rhomb (64QAM) marks placed over the modulation capacity curves relate the effective code rate to the BICM SNR threshold for each I MCS format (identified by the numbers near the marks). Those points are the absolute capacity limits of the different formats from an information theory point of view. Fig. 8. Effects of code block fragmentation on the BLER curves in AWGN channel The square (QPSK), asterisk (16QAM) and star (64QAM) marks relate the net throughput (net transferred information rate in bits/llr) to the SNR threshold for a BLER of 10% (SNR 10% ). The units are bits/llr because here by channel use we understand the reception of a single LLR carrying intrinsic information of a given coded bit. Notice that for I MCS formats 0 to 4, which imply bit repetition, a single coded bit may give rise to the reception of several LLRs. The net throughput at SNR 10% is 90% of the effective code rate of the I MCS format. In Fig. 9, for the I MCS formats 0 to 9 (QPSK), the plus marks denote the net throughput in bits/coded bit at SNR 10%. Here by channel use we understand the reception of a combined LLR at the input of the turbo decoder, so that any coded bit maps to a single combined LLR. If there is bit repetition then the combined 7

9 LLR is the addition of the individual LLRs belonging to the same coded bit. For I MCS formats 5 to 9 there is no bit repetition (any coded bit maps to a single LLR) and the net throughput in bits/coded bit at SNR 10% is the same as the net throughput in bits/llr. For I MCS formats 1 to 4 the net throughput saturates at 0.3 [bits/coded bit]. This is because for an effective code rate equal r (r < 1/3) we are receiving (on average) /r LLRs per coded bit, and since each LLR supports 0.9 r [bits/llr] of average information, the net throughput is 0.9 r /r = 0.3 [bits/coded bit]. I MCS =0 is special case, since the LLR combiner (placed before the turbo decoder) will remove, from the first constituent block, the 8 LLRs carrying the intrinsic information of the filling bits and substitute them by ideal LLRs; and it will, additionally, insert 8 ideal LLRs at the beginning of the systematic block. Since each of the ideal LLRs supplies exactly 1 bit of information and the effective code rate for I MCS =0 is 0.17, we can say that 8 ideal LLRs are equivalent to 8/0.17=63 LLRs from the channel. Then, from the point of view of the turbo decoder, the received information is equivalent to =370 LLRs for a code block of 40 systematic bits, or [LLRs/coded bit] (notice that 40 systematic bits are equivalent to 10 coded bits). Then the throughput is [bits/llr] [LLRs/coded bit] = 0.35 [bits/coded bit]. Here we count the filling bits as useful throughput (although they are not) to be able to compare with the experimentally measured capacity of the discrete memoryless channel (DMC) (Fig. 10). We can see that, thanks to the energy gain obtained through bit repetition, the E b /N 0 for I MCS 0 to 3 is kept at the right value to achieve BLER=10% even though the SNR has been drastically reduced. For I MCS =9, for example, the coordinates of the circle mark are (1.93, 0.64) since the BICM threshold is 1.93 db and the effective code rate is The coordinates of the plus and square marks are (3.4, 0.576) because for the SNR 10% of 3.4 db the net throughput is =0.576 [bits/llr]. Fig. 9. Capacity and net throughput for all I MCS (1 RB) in AWGN channel At our simulator we are able to measure the mutual information between the transmitted bits and the received (combined) LLRs. This is achieved by characterizing the channel between the transmitted coded bits and their corresponding LLRs as a DMC (see Fig. 10). 8

10 b b {0,1} Channel z=llr(b) Fig. 10. Equivalent DMC between a transmitted coded bit and its LLR The capacity for that channel is the input-output mutual information (MI), given by (see [4]): fzb ( z b) 1 MI( b, z) = Eb, z log log f zb i( z b i) dz = = = f ( ) i 0,1 fzb i ( z b i ) z z = = = 1+ fzb = i( z b= i) () where f z (z) is the p.d.f. of the LLRs. By numerically computing expression () we obtain the cross marks in Fig. 9, which represent the measured MI [bits/coded bit] at SNR 10% for I MCS formats 0 to 9 (QPSK). For I MCS 5 to 9 there is no bit repetition, so the measured MI is equal to the modulation capacity (as expected). For I MCS 0 to 4 the capacity is above the net throughput by an amount approximately equal to the gap present for I MCS 5 to 9. We can interpret that gap as the excess of information needed to keep BLER=10% with a short codeword. Due to the incremental redundancy (IR) HARQ scheme, for every new rv received the LLRs are combined before trying to decode the codeword. The combination of LLRs, which is equivalent to MRC processing, reduces the number of punctured bits and thus reduces the effective code rate as seen by the decoder. It also increases the reliability for bits that have been received more than once. Both effects lead to a remarkable increase in the probability of correct decoding. Rate matching is responsible for creating a well designed puncturing pattern of the mother code. It must be implemented exactly as the standard specifies, since the decoding process is sensitive to the puncturing pattern. Interleaving is always necessary, even for simulations assuming AWGN channel, since the multilevel modulations (mainly 64QAM) have unequal error protection of the coded bits and create a periodic pattern in the reliability of the received LLRs that must be broken by the interleaving (this is the principle of BICM). The implementation of the rate matching and HARQ are one of the most sensitive pieces within the link level simulator, since small differences in the interpretation of the standards can lead to different results. The rate matching is performed, on the basis of a single code block, by selecting the coded bits to transmit from a buffer which is circularly addressed, see [4]. For a code block of size N (40 N 6144), the size of the circular buffer is 3 N+1. In case of fragmentation of the TB we need to implement as many circular buffers as code blocks. The circular buffer is filled with the coded bits already interleaved; then it is read by sectors and sent to the channel in an ordered way. For every TTI of 1 ms a single sector from each circular buffer is transmitted. The size of the sector depends on the code rate (as specified by the I MCS format). For a code rate of 1/3 the sector size equals the full buffer, while for a rate below 1/3 the sector encompasses more than a single turn, which means that some of the coded bits are transmitted more than once. In order to implement IR, every HARQ rv starts reading the circular buffer at a different offset. At the receiver side, the programming of the turbo decoder is also sensitive to different implementations. The turbo internal interleaving and de-interleaving has been changed with respect to the 3G (UTRA) specifications. There are different alternatives for soft-input soft-output decoders for the two convolutional codes that are the constituents of the turbo code. Our simulator uses Maximum a Posteriori Probability (MAP) algorithm and a maximum of 8 decoding iterations. The number of iterations has an impact on the BLER. 9

11 The computation of the bit level LLRs also may have an impact on the BLER. The definition of the bit level LLR is: Prob( 0 ) Prob( 0) ln b = y ln y b = Λ= = Prob( b= 1 y) Prob( y b= 1) equally probable bits (3) where y is the received variable of decision, i.e., the complex sample after MIMO equalisation. The computation of expression (3) depends on the modulation and on the noise statistics. Assuming Gaussian noise we get: Λ= ln iy : i S0 jy : j S1 exp exp y y σ i y y σ j (4) where S 0 is the set of modulation states for which the transmitted bit is at logical zero, S 1 is the set for which the transmitted bit is at logical one and σ is the complex noise variance. Since expression (4) is too complex to evaluate, in particular for high order modulations, it is usually approximated by: y y 0 exp σ 1 Λ ln = y y σ 1 exp σ ( y y1 y y0 ) (5) where y 0 (resp. y 1 ) is the modulation state for which the transmitted bit is at logical zero (resp. logical one) that falls nearest to y. The simplification is based on the max-log approximation, i.e.: a b ln( e + e) max( ab, ). If the noise is not Gaussian, which can happen for example when there is residual interference after equalization, it is possible to take the noise statistics into account in the computation of the LLRs. In Fig. we compare the statistics of the bit level LLRs for 64QAM when computed based on expressions (4) or (5). Although the LLRs can take big absolute values (mainly for high SNR), in expression () we can see that the region of LLR values for which the p.d.f. s do not overlap contributes to the MI as an integral value. This means that we can compress the dynamic range of the LLRs by limiting their absolute value to the highest value for which the p.d.f. s for bit=0 and bit=1 still overlap. In this way the p.d.f for the non overlapping region becomes a Dirac delta function that contains the same area as the original p.d.f., and the MI of the channel is preserved. Given the range of SNRs of interest, in our simulator we limit the LLRs to an absolute value not greater than 10. In order to summarize, in Table we list the set of features that could be shared in order to align the TB processing of different link level simulators. 10

12 Fig.. Comparison among exact and approximated LLRs for 64QAM and SNR=7.16 db Table. Summary of main aspects of the TB processing chain that may need alignment Number of simulated codewords: 1, Rate Matching: Computation methodology of the PDSCH capacity, the PDSCH payload size, the TB segmentation and the filling bits Table of the effective code rate as a function of the MCS index Circular buffer implementation: size and starting offset for the different rv s Puncturing patterns created after rate matching for the different code rates and redundancy versions Bit repetition patterns after LLR combination for the different rv s Turbo decoder implementation details: LLR computation methodology constituent decoders and iterations Trellis finalisation bits Reference throughput and BLER curves in AWGN channel for the different MCS indexes and for the 4 HARQ rv s 5. LTE / LTE-A Transmission modes MIMO techniques can be configured in different ways to provide a combination of gains in throughput, diversity and beamforming. From the point of view of the channel state information (CSI) the MIMO system can be open-loop (OL) or closed-loop (CL). While OL MIMO systems only have CSI at the receiver, CL MIMO systems also know the CSI at the transmitter and this can be used to improve the throughput and reliability. Single-user MIMO (SU-MIMO) schemes dedicate all spatial layers to one user, while Multi-user MIMO (MU-MIMO) schemes allow multiple users to be co-scheduled on the same time and frequency resources. The set of MIMO schemes to be applied between the enb and the UE for

13 both DL and UL are defined by the LTE/LTE-A Transmission Modes (TM). Release 8 supports 7 TM and allows a maximum of 4 Tx antennas for DL and 1 Tx antenna for UL. The baseline MIMO antenna configuration (N Tx x N Rx ) for Release 8 is x for DL and 1x for UL. LTE-A has increased the number of antennas to 8 Tx antennas for DL, 4 Tx antenna for UL and up to 8 Rx antennas. LTE-A (Release ) supports 10 TM. Therefore MIMO configurations of 8x8 for DL and 4x4 for UL are allowed in LTE-A, and this is one of the key aspects to meet the LTE-A requirements. The 7 TM defined in Release 8 are: TM 1: Single antenna port, port 0 This mode corresponds to a 1x SIMO configuration, since it only uses 1 Tx antenna and Rx antennas. MRC is applied at the receiver to obtain receive diversity gain. TM : Transmit diversity This is the OL Transmit diversity mode for a single user based on space-frequency block coding (SFBC) techniques with Alamouti Code at the transmitter complemented with frequency-shift time diversity (FSTD) when 4 transmit antennas are used. It supports or 4 Tx antennas, up to 4 spatial layers and only 1 codeword. MRC combining is used at the receiver. TM 3: Large-delay CDD This is the OL MIMO SM mode for a single user, and supports or 4 Tx antennas, up to 4 spatial layers and up to codewords. At the receiver side there are at least Rx antennas, so x, 4x or 4x4 MIMO configurations are possible. The LTE specifications include a fixed large delay CDD precoding at the enb that consists in transmitting the same OFDM symbols on the same set of subcarriers from multiple antennas with a different delay on each antenna. This creates an artificial multipath that translates into additional frequency diversity, which is then exploited by the turbo code. TM 4: Closed-loop spatial multiplexing This is the CL MIMO SM mode for a single user and supports or 4 Tx antennas, up to 4 spatial layers and 1 or codewords. The same MIMO configurations of TM3 are possible, i.e., x, 4x or 4x4. This mode requires a precoding of the spatial layers based on a codebook, defined by the standard, which is known at both enb and UE. The UE estimates the MIMO channel conditions and feeds back the precoding matrix index (PMI) to the enb. TM 5: MU-MIMO This is the CL MIMO SM mode for multiple users. In the UL the number of simultaneous UEs on the same frequency and time resources is limited by the number of Rx antennas at the enb, while in the DL it depends on the precoding techniques available at the enb. With Release 9 up to 4 UE are supported in the DL ( UE in Release 8). TM 6: Closed-loop spatial multiplexing using a single layer This mode is a particular case of TM 4 when only one spatial layer is used (rank=1). The UE feeds back to the enb the best PMI for capacity maximization. This precoding results in certain beamforming gain but, due to the use of a restricted codebook, it does not correspond to the UEspecific beamforming where the beam is directed to the selected UE. TM 7: Beamforming. Single antenna port, UE-specific RS (antenna port 5) This mode is the UE-specific beamforming mode where only one layer is transmitted to one UE and the enb uses a virtual antenna (port 5) to direct the antenna beam to the UE. LTE Release 9 added the TM 8: TM 8: Dual layer beamforming (antenna port 7 and 8) This mode is similar to TM 7 but it specifies a single or a dual layer transmission on antenna ports 7 and 8; therefore this mode allows the enb to dedicate the two layers to one UE (singleuser) or two UEs, one layer per user. LTE-A Release 10 added the TM 9: 1

14 TM 9: Closed-loop SU/MU-MIMO (antenna ports 7 to 14, UE-specific and CSI Reference Signals) This mode allows MIMO configurations up to 8x8, dynamic SU/MU-MIMO switching and supports the new reference signals introduced by LTE-A: the UE-specific reference signals (DM-RS) for data demodulation and the CSI reference signals (CSI-RS) for downlink channel state information measurement by the UE. The new DM-RS allow non-codebook based precoding for the CL SU/MU-MIMO. The current state of the LTE-A specifications is the Release and it has introduced the TM 10 which is very similar to TM 9, except that TM 10 allows a UE to be configured with one or more CSI processes per serving cell. Table 3 summarizes the LTE/LTE-A Transmission Modes for PDSCH. According to TS the Tx diversity scheme is available in each TM a fallback; so the MIMO configuration can switch from SM to Tx diversity or vice-versa depending on the channel conditions. Table 3. LTE/LTE-A Transmission Modes for PDSCH TM Transmission scheme of PDSCH Number of Antennas MIMO Operation SU/MU MIMO Releases 1 Single-antenna port, port 0. 1 Open-Loop SU-MIMO Transmit diversity, 4 Open-Loop SU-MIMO 3 Open-loop spatial multiplexing (with large delay CDD Precoding). 4 Closed-loop spatial multiplexing, 4 5 Multi-user MIMO, 4 6 Closed-loop spatial multiplexing with a single transmission layer 7 Beamforming, Single-antenna port, port 5., 4 Open-Loop SU-MIMO, 4 1,, 4 8 Dual layer beamforming (antenna ports 7 and 8), Closed-loop SU/MU-MIMO (antenna ports 7 to 14, UE-specific and CSI Reference Signals) Closed-loop SU/MU-MIMO (antenna ports 7 to 14, UE-specific and CSI Reference Signals). With one or more CSI processes., 4, 8, 4, 8 Closed- Loop Closed- Loop Closed- Loop Closed- Loop Closed- Loop Closed- Loop Closed- Loop SU-MIMO MU-MIMO SU-MIMO SU-MIMO SU/MU MIMO SU/MU MIMO SU/MU MIMO 8, 9, 10, 8, 9, 10, 8, 9, 10, 8, 9, 10, 8, 9, 10, 8, 9, 10, 8, 9, 10, 9, 10, 10, LTE introduced only one transmission mode for UL, so the TM 1 for PUSCH was the single-antenna port scheme with 1 Tx antenna at the UE and at least Rx antennas at the enb (to enable Rx diversity). Therefore, the TM 1 does not support SM for a SU-MIMO operation, but MU-MIMO can be used in the UL to enhance system capacity. LTE-A has introduced the TM for PUSCH, which consists on a CL SM MIMO scheme. With this mode up to 4 layers can be transmitted from the same UE and a precoding is applied before transmission to adapt to channel conditions. In the UL there is no Tx diversity fallback mode, so if CL SM is not possible then the system returns to the single-antenna port mode. Table 4 summarizes the transmission modes for PUSCH. 6. PHY channel processing The PHY channel processing is the lowest layer in the link level simulator. It takes as input the coded bits of the codewords (1 or codewords) after rate matching and interleaving and processes them according to the selected transmission mode. On the receiving side the LLRs of the coded bits are delivered to the deinterleaver and LLR combiner on the TB processing chain. Notice that if Successive Interference 13

15 Cancellation (SIC) techniques are applied the division between PHY channel processing and TB processing is not so sharp, since the already decoded code blocks are sent back to the PHY processing and used to improve the MIMO equalisation. Table 4. LTE/LTE-A Transmission Modes for PUSCH TM Transmission scheme of PUSCH Number of Antennas MIMO Operation SU/MU MIMO Releases 1 Single-antenna port scheme 1 Open-Loop MU-MIMO 8, 9, 10, Closed-loop spatial multiplexing scheme, 4 Closed-Loop SU/MU MIMO 10, There are some high level aspects that must be initially specified, like the LTE Release (LTE or LTE-A), the simulation of FDD/TDD, the simulation of DL or UL and the simulation of a unicast carrier or a broadcast (e-mbms) carrier. Next high level issue is the simulated LTE/LTE-A transmission mode. In addition to the LTE-A transmission mode, which conditions the full PHY configuration, important parameters to settle down are: simulated carrier frequency and simulated bandwidth. Those parameters have an impact on the channel model that can be applied and on the PHY layer capacity, which has an influence on the size of the code blocks that can be applied. In [9] there is a proposal for E-UTRA operating bands for FDD and TDD, as well as deployment scenarios for feasibility study. The RBs are allocated in groups of different sizes depending of the whole system bandwidth. The RBs can be allocated contiguously or scattered in the whole system bandwidth. In the UL, due to the use of SC-FDMA, the RBs are allocated contiguously but a frequency hopping pattern may be applied, on a TTI basis, to increase frequency diversity. The OFDM modulation accepts several configurations: the subcarrier spacing can be 15 khz or 7.5 khz, this last for e-mbms subcarriers and single frequency networks (SFN). The normal cyclic prefix duration is about 5 µs but it is also possible to use an extended cyclic prefix of 17 µs (for rural cells) and 33 µs for e-mbms. With the normal cyclic prefix there are 14 OFDM symbols per subframe (1 ms), while for the extended cyclic prefix there are only 1 OFDM symbols per subframe. The available PHY capacity per subframe depends on the room reserved for control channels and RS (pilots). The simulation of the OFDM modulation may be based on DFT or FFT. The advantage of using DFT over FFT is that multipath delays (at the channel simulator) can take any value, while FFT defines a sampling period that must be respected and this forces the channel delays to be integer multiples of the FFT sampling period. Depending on the type of channel simulator, and if perfect time and frequency synchronisation are assumed, it is possible to completely skip the IFFT/FFT stages in the DL simulator, since the channel can be simulated in the frequency domain and the receiver noise can also be added to the received signal samples in the frequency domain. The receiver noise becomes correlated after the FFT stage, but since OFDM spectrum accomplishes the Nyquist criterion in the frequency domain, noise samples taken at the subcarrier frequencies become uncorrelated. If IFFT/FFT stages are not simulated then we are assuming that the channel remains constant for the full duration of the OFDM symbols, so that channel variations due to Doppler effect are updated in quantized time steps equivalent to the OFDM symbol duration. Another aspect related to OFDM is the possible simulation of PAPR reduction techniques. The simulation of the MIMO multipath channel is one of the key issues in the simulation of the PHY channel. Depending on the desired results and on the simulated LTE transmission modes the channel simulator may introduce different degrees of complexity. A classical way to simulate the MIMO matrix channel is to follow the procedure outlined in [10], [], which is a purely stochastic method. The model consists of a power delay profile, with several taps, which depends on the propagation environment (EPA, EVA, ETU). For each propagation path a complex random matrix, with i.i.d. elements is generated. Antenna correlation is introduced by applying Cholesky factorisation to the original i.i.d. elements based 14

16 on a set of matrices (from [10]) to model low, medium or high correlation. Time variations are simulated by using a classical Doppler spectrum (Jakes) with a maximum Doppler frequency which also depends on the propagation environment. A more complex method is the one proposed in [1], [13], [14]. This method takes into account the per-path power azimuth spectrum at the enb and at the UE, so it models the geometry of the scattering in a stochastic way. Given that the TTI in LTE is 1 ms, in slow mobility scenarios it is possible to make the hypothesis that the channel is constant during the transmission of a complete TB. If we also assume that the UE is not scheduled persistently, then the channel realisations in consecutive transmissions can be considered independent. In this case it is not necessary to model the Doppler variations of the channel, and the simulator can generate independent channel snapshots for each TB. Obviously, this approach is not valid for continuous transmission like, for example, in a broadcast (e-mbms) carrier. A topic related to the channel simulation is channel estimation. Here the options are: a) channel estimation is assumed ideal, b) pilot symbols and real channel estimation methods are simulated or, c) channel estimation error is modelled as an additive Gaussian noise power. The options b) and c) are more realistic, mainly for high mobility scenarios. Option c) requires training of a model for channel estimation errors, [15]. There are several options for channel estimation that need to be coordinated for simulator calibration, like the type of time/frequency domain interpolations and the use (or not) of Wiener filtering. The power boost (if any) applied to pilots should also be considered. The pilots can be the cell-specific Common Reference Signals (CRS) or UE Specific Reference Signals (UE RS), []. UE RS are used for the transmission modes where the enb applies a UE specific MIMO precoding matrix. In this case the UE cannot estimate the channel based on the CRS, then either the precoding is based on a codebook and signalled to the UE, or the enb sends UE RS that use the same precoding that has been applied to the data symbols. In this last case the precoding matrix is not restricted to any codebook. The MIMO processing at the PHY channel is linked to the simulated transmission mode. It is on the receiver side and for the spatial multiplexing (SM) modes where there are many options for the MIMO processing. The BLER and throughput performances of the link level can be heavily influenced by the number of antennas and the implemented MIMO processing. On the transmitter side the mapping from codewords to layers and antenna ports must be specified, as well as AMC and power allocation to the different MIMO layers. For closed loop SM, the method for selection of the precoding matrix (either within a codebook or unrestricted) must be specified. For single user SM and for multi-user MIMO modes, the MIMO equalisation technique (ZF, MMSE, ML, SIC, etc.) must be specified. SIC can be applied on a layer basis or on a codeword basis. In this case the CRC of the codeword is verified to ensure that errors are not fed back by the SIC processing. Table 5 summarizes the aspects that we have identified that should be put in common for simulator calibration with respect to the PHY channel processing. Table 5. Summary of main aspects of the PHY channel simulation that may need alignment High level aspects: Supported LTE Releases FDD / TDD UL / DL Unicast / e-mbms Supported LTE transmission modes Carrier frequency and simulated bandwidth (contiguous or scattered RBs?) OFDM aspects: Subcarrier spacing and cyclic prefix duration Simulated sampling rate Number of FFT points Simulation of time and frequency synchronisation errors PAPR reduction techniques 15

17 Simulation of the MIMO multipath channel: Purely stochastic (3GPP TS ) Geometry-based stochastic (3GPP TR 5.996, ITU-R M.135) Traces from measurement campaign Others Simulation of Doppler spectrum or channel snapshots Channel estimation: Ideal channel estimation Pilots are simulated: linear interpolation, Wiener filter (SNR and channel correlation assumed known/unknown), pilots power boost Channel estimation error is simulated as additive Gaussian noise MIMO processing: Number of antennas at Tx and Rx Number of spatial layers Tx diversity modes (Alamouti) AMC and power allocation to MIMO layers Spatial multiplexing modes: o Mapping of codewords to layers and antenna ports o Open loop SM with CDD o Closed loop SM: selection of precoding matrix (codebook or unrestricted) o SU-MIMO equalisation: ZF, MMSE, ML, SIC, etc. o MU-MIMO precoding and equalisation Other issues: Simulation of HARQ retransmissions at PHY channel (delays and channel variation between rv s) CQI reporting and effects of delay CoMP, Relays, Carrier aggregation, etc. 7. Link abstraction Link abstraction techniques aim at obtaining look-up tables to predict the link performance (BLER) for multistate channels, [5]. A multistate channel arises when the received LLRs, within a given codeword, show very different reliabilities. This is typically due to frequency selective fading, which may show important variations throughout the OFDM subcarriers. The LLR combination that happens before decoding when HARQ is in use also creates a multistate channel, and even the unequal error protection of high order modulations can be interpreted as a multistate channel. Link abstraction techniques take as input the SNR of each subcarrier (or group of subcarriers) and obtain a single scalar value, called the Effective SNR that can predict the BLER of the link. The reference BLER curves (BLER for the different AMC formats in AWGN channel) play an important role in this methodology. Depending of the desired accuracy, link abstraction techniques may become complex. One popular choice is Exponential Effective SNR Metric (EESM), which derives from the union bound for the BLER performances of convolutional codes, [6], [7]. The advantage of EESM is that the weighting function that is used for averaging the SNR of the subcarriers has a closed form, but EESM requires training, i.e. to obtain good BLER predictions it is necessary to previously obtain a set of parameters which are dependent on the AMC format and on the transmission mode. Another possibility is the Mutual Information Effective SNR Metric (MIESM), [5], [8], which outperforms EESM in terms of BLER prediction accuracy and does not require training. Table 6 summarizes the aspects that may need coordination among the different simulators with respect to link abstraction. 16

18 Table 6. Summary of main aspects of the Link Abstraction procedures that may need alignment Supported Link Abstraction methodologies: EESM, MIESM... Calibration methods Simulation of HARQ at system level Generated Look-up tables: BLER vs. Effective SNR for the different CQI s (or MCS indexes) Tables of calibration parameters for the different CQI s (or MCS indexes) CONCLUSIONS In this paper we have reviewed the main aspects to consider when designing a LTE link level simulator with the objective to point out the parameters and configuration options that must be agreed in order to obtain results that can be compared. It has been shown that there are many options to consider, an so the emphasis and the simulation efforts should be put on the most interesting transmission modes from a research point of view. The future work is to foster collaboration among the interested institutions in order to interchange simulator results with the aim of achieving a common reference for LTE-A link level simulators calibration and a set of link abstraction look-up tables to provide support to the system level simulations. ACKNOWLEDGMENT This work has been supported by the Spanish Ministry of Science under the project TEC C0-01. REFERENCES [1] 3GPP TS 36.1, E-UTRA Multiplexing and Channel Coding [] 3GPP TS 36., E-UTRA Physical Channels and Modulation [3] 3GPP TS 36.13, E-UTRA Physical Layer Procedures [4] J. Olmos, A. Serra, S. Ruiz and I. Latif, On the Use of Mutual Information at Bit Level for Accurate Link Abstraction in LTE with Incremental Redundancy H-ARQ, COST IC1004 TD(1)05046, Bristol, UK, 4-6 September, 01 [5] K. Brueninghaus, D. Astdlyt, T. Silzert, S. Visuri, A. Alexiou, S. Karger, G. Seraji, Link Performance Models for System Level Simulations of Broadband Radio Access Systems, PIMRC 005 [6] 3GPP-C , Effective-SNR Mapping for Modeling Frame Error Rates in Multiplestate Channels, Ericsson [7] J. Olmos, A. Serra, S. Ruiz, M. García-Lozano, D. Gonzalez, Exponential Effective SIR Metric for LTE Downlink, IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, PIMRC 009 [8] J. Olmos, Silvia Ruiz, M. García-Lozano, D. Martín-Sacristán, Link Abstraction Models Based on Mutual Information for LTE Downlink, COST 100 TD(10)05, Aalborg, Denmark, -4 June, 010 [9] 3GPP TR , LTE-Advanced feasibility studies in RAN WG4 [10] 3GPP TS , E-UTRA User Equipment (UE) radio transmission and reception [] 3GPP TSG RAN WG4, R , Radio Propagation Modelling for E-UTRA performance requirement definition [1] 3GPP TR 5.996, Spatial channel model for MIMO simulations [13] 3GPP TR , E-UTRA Further advancements for E-UTRA physical layer aspects [14] ITU-R M.135, Guidelines for evaluation of radio interface technologies for IMT-Advanced [15] A. Serra, J. Olmos, M. Lema, Modelling Channel Estimation Error in LTE Link Level Simulations, COST IC1004 TD(1)03067, Barcelona, Spain, 8-10 February, 01 17