Interference-Aware Receivers for LTE SU-MIMO in OAI Elena Lukashova, Florian Kaltenberger, Raymond Knopp Communication Systems Dep., EURECOM April, 2017 1 / 26
MIMO in OAI OAI has been used intensively for MIMO research Advanced receivers for Multi-user MIMO (transmission mode 5) 1 Channel sounding campaings 2 Massive MIMO beamforming (transmission mode 7) using reciprocity 3 But basic SU-MIMO was long neglected This work describes the receiver architectures that have been implemented for 2x2 TM3 (open-loop) and TM4 (closed loop) 1 Wagner and Kaltenberger, Interference-Aware Receiver Design for MU-MIMO in LTE: Real-Time Performance Measurements. 2 Kaltenberger et al., Broadband Wireless Channel Measurements for High Speed Trains. 3 Jiang et al., MIMO-TDD Reciprocity and Hardware Imbalances: Experimental Results. 2 / 26
Receiver Architectures ML receivers are optimum, but have high computational complexity. Linear receivers have poor perfomance (especially in fixed point) Reduced complexity Maximum-Likelihood (R-ML) receivers based on Interference-Aware (IA) soft bit LLR metrics 4. Parallel Interference Aware (PIA) receiver: suboptimal IA detection of first stream + successive interference cancellation (IA-SIC): does not introduce any informational loss compared to ML MIMO receiver 4 Ghaffar and Knopp, Low Complexity Metrics for BICM SISO and MIMO Systems. 3 / 26
Signal model for TM3 and TM4 The received signal vector y l C 2 1 for the l-th subcarrier seen by the UE is given by: y l = H l P l x l + n l, [ ] K 1 e jφ H l = + K + 1 e jφ 1 } {{ } LOS component l = 1, 2..., L, where 1 K + 1 H l, }{{} NLOS component φ is a phase-shift and K is the Rician K-factor, P is a precoding matrix. y = H eff x + n, H eff = [h eff0 h eff1 ]. 4 / 26
Precoding in TM3 and TM4 TM3 Cyclic Delay Diversity applies a different phase delay for each modulation symbol i: For 2 transmission layers, W(i) = 1 2 I, D(i) = P CDD (i) = W(i)D(i)U. [ 1 0 0 e πi ], U = 1 [ ] 1 1 2 1 e jπ. TM4 precoders for 2 transmission layers: choose P from the following options { [ ] 1 1 1 P, 1 [ ]} 1 1. 2 1 1 2 j j 5 / 26
R-ML Parallel Interference-Aware Receiver Both codewords undergo matched filtering and are handled in the same manner using interference-aware low-complexity Log-Likelihood Ratio (LLR) metrics: LLR 0 M0_M1_llr LLR 1 M1_M0_llr 6 / 26
R-ML Successive Interference Canceling Receiver After Matched filter, CW 0 is handled identical to the PIA detection, while CW 1 receives interference-free treatment: LLR 0 M0_M1_llr LLR 1 M1_llr 7 / 26
Mutual Information Potential of PIA and SIC receivers Based on the MI chain rule, the SIC receiver is informationally lossless compared to the joint ML detection. I ML = I SIC = I (X 0; Y MF H eff ) }{{} + I (X 1; Y MF X 0, H eff ) }{{} MI for the first CW MI for the second CW using SIC I ML I PIA0 +I PIA1 = I (X 0; Y MF H eff ) }{{} + I (X 1; Y MF H eff ) }{{} MI for the first CW MI for the second CW using PIA 8 / 26
Mutual Information Potential MI,[bit/symb] 12 10 8 6 4 2 SIC 64 64QAM PIA 64 64QAM SIC 16 64QAM PIA 16 64QAM SIC 16 16QAM PIA 16 16QAM 5 10 15 20 25 30 SNR, [db] Our SIC receiver possesses higher MI potential compared to PIA. The main gains come from 64 64QAM constellations. 9 / 26
OAI Simulation Parameters Fading environment: 1-tap and 8-tap Rayleigh and Rician channels, EPA model with moderate (EPAM) and high (EPAH) correlation. Rician channel: AoA α { π 4, π 8 } radians and K-factor of 9.5 db; Throughput simulations: 3000 packets with 1 PDCCH symbol. Bandwidth 5MHz. 10 MCS 0 28, MCS 0 MCS 1 28. Target BLER 10 2. Perfect Channel Estimation. 10 / 26
Imperical Performance: Rician channel 35 Total Throughput, [Mbit/s] 30 25 20 15 10 5 SIC Rice AoA π 4 PIA Rice AoA π 4 SIC Rice AoA π 8 PIA Rice AoA π 8 4Mbit/s 5 10 15 20 25 30 SNR, [db] Our SIC receiver demonstrates 4Mbit/s throughput gain in Rician channel compared to PIA. 11 / 26
Computational Effort SIC16-16 PIA16-16 SIC64-64 PIA64-64 Trial processing duration (µs) 1,488 1,197 1,104 662 369362 181 253 79 148 0 0 MF and LLR SIC block Full rec For the high modulation order the SIC receiver is 25% more time efficient than PIA receiver, since the SIC block that takes less time than the 64 64QAM IA LLR metric for the CW 1. 12 / 26
Link Adaptation in TM3 and TM4 Channel State Information: the Rank Indicator (RI) the Precoder Matrix Indicator (PMI) the Channel Quality Indicator (CQI) To tune the transmission in the way to guarantee high levels of throughput under certain reliability constrains, the UE estimates: the RI and the CQI for TM3 the RI, the PMI and the CQI for TM4 13 / 26
Rank Indicator Estimation RI 2: CDD in TM3 and CLSM in TM4 RI 1: Alamouti precoding. Proposed criteria for 2 2 MIMO: Condition Number κ(h). κ(h) = H F H 1 F, If κ(h) is less than a certain threshold T, the UE reports RI 2, otherwise RI = 1. 14 / 26
Rank Indicator Estimation CLSM transmissions, [%] 100 80 60 40 20 0 8-tap Rayleigh 5 10 15 20 25 100 80 60 40 20 0 EPAH 5 10 15 20 25 T, [db] T, [db] With the increase of the correlation level, the amount of the simultaneous transmission of the two transport blocks given the same values of threshold T decreases. Theoretically we might expect that in the highly correlated channels rank adaptation is important. 15 / 26
Rank Indicator Estimation SIC EPAH PIA EPAH Throughput, [Mbps] 10 8 6 4 2 10 8 6 4 2 5 10 15 20 25 30 5 10 15 20 25 30 T = 5dB T = 10dB T = 15dB T = 20dB T = 25dB no RA SNR, [db] SNR, [db] The throughput values are the highest, if the enodeb always sends two transport blocks regardless of the channel conditions. Thus, the R-ML PIA and SIC detection demonstrate high throughput when no rank adaptation is applied. 16 / 26
Precoder Estimation for TM4 For TM4 with 2 active antenna ports the 3GPP standard defines 2 precoding options: { [ ] 1 1 1 P, 1 [ ]} 1 1. 2 1 1 2 j j The two most common metrics: maximum mutual information criteria. It involves a time-consuming computation of MI or heavy look-up tables. maximum SNR criteria. Intuitive and light implementation solution is to min BLER of CW 0 to increase probability of decoding for CW 1 5. 5 Ghaffar and Knopp, Making multiuser MIMO work for LTE. 17 / 26
Comparison of the max MI and max SNR Criterion 12 MI for 2 CWs,[bit/symb] 10 8 6 4 2 16-16 MI based 16-64 MI based 64-64 MI based 16-16 SNR based 16-64 SNR based 64-64 SNR based 5 10 15 20 25 30 35 SNR, [db] The MI-based criterion outperforms the SNR-based computation only when CW 0 is mapped on 16QAM, and CW 1 belongs to 64QAM. This gap vanishes when both codewords belong to the same constellation. 18 / 26
Current status in OAI Our PIA and SIC receivers are implemented in dlsim simulator. TM3 has been tested with PIA receiver in real-time, SIC and TM4 to come soon. RI and PMI estimation are supported, CQI to come soon. HARQ for both SU-MIMO with both PIA and SIC is supported. TM3 enb in test phase. Work on 5G receivers has begun. 19 / 26
5G MIMO Schemes MIMO CLSM schemes for 5G 6 : 1 CW for 1 to 4-layer transmission 2 CWs for 5 to 8-layer transmission. SIC can no longer be used in the same way! Possible solutions for 2 layers PIA can still be used SIC can be used, but without CRC risk of error propagation 6 3GPP, Study on New Radio Access Technology.Physical Layer Aspects. 20 / 26
Possible solutions for 4 layers Extend LLR metrics to take into account 4 interfering symbols; Reduce 4 4 channel such that the interference is reduced to 2 2 matrix so that the existing LLR metrics can be used. Apply practically feasible Block QR decomposition: 7 HP = Q 0Q 1R, [ ] [ ] [ ] Q00 Q 01 I 0 R00 R 01 y = x + n = Q Q 10 Q 11 0 Q1 0 R 0Q 1Rx + n 11 Signal Transformation [ ] Q T R00 R 01 0 y = x + Q 0 Q1R T 0 n, 11 Apply PIA or SIC to the 2x2 sub-blocks. 7 Thomas et al., Detection using block QR decomposition for MIMO HetNets. 21 / 26
Thank You! http://www.openairinterface.org/ 22 / 26
Interference-Aware soft bit LLR metrics The distance metric of ML detector D(y H eff x) = arg min y h eff0 x 0 h eff1 x 1 2 x 2 M 0,M 1 LLR of x 0 using the MaxLog approximation of ML receiver of IA detection: λ (y, H, x) = max { D(y H eff x)} x 2 M 0,M 1 The IA detection allows to reduce the complexity from dual stream 2 M0+M1 to single stream 2 M0 thanks to defining for each symbol x 0 a single optimal interfering symbol x 1. The exact metrics for QPSK, 16QAM, 64QAM are derived by Ghaffar and Knopp 8. 8 Ghaffar and Knopp, Low Complexity Metrics for BICM SISO and MIMO Systems. 23 / 26
Interference-Aware soft bit LLR metrics λ = max x 2 M 0,M 1 { h eff0 2 x 0 2 h eff1 2 x 1 2 + 2 [ R(y MF0 )R(x 0 ) + I(y MF0 )I(x 0 ) ] η 0 = R(ρ)R(x 0 ) + I(ρ)I(x 0 ) R(y MF1 ), η 1 = R(ρ)I(x 0 ) I(ρ)R(x 0 ) I(y MF1 ) 2 η 0 R(x 0 ) 2 η 1 R(x 1 ) }, To maximize λ, the signs of the real and imaginary parts of the interfering symbol x 1 should be opposite to η 0 and η 1 respectively, and R(x 1 ) opt = η 0 h eff1 2, I(x 1 ) opt = η 1 h eff1 2. 24 / 26
Optimal MCS in Rayleigh channel MCS 28 26 24 22 20 18 16 14 12 SIC Rayleigh PIA Rayleigh 5 10 15 20 25 30 SNR, [db] 5 10 15 20 25 30 SNR, [db] MCS 0 MCS 1 The SIC receiver supports higher MCS for CW 1 thanks to the interference-free detection of the second stream. 25 / 26
Precoder Estimation for TM4 We propose to evaluate the correlation coefficient ρ 10 = h H eff1 h eff0. P = 1 2 1 2 [ 1 1 1 1 [ 1 1 j j ] ], for R(ρ 10 ) I(ρ 10 ), for R(ρ 10 ) < I(ρ 10 ) 26 / 26