MU-MIMO scheme performance evaluations using measured channels in specific environments Christoph Mecklenbräuker with contributions from Giulio Coluccia, Giorgio Taricco, Christian Mehlführer, and Sebastian Caban www.ist-mascot.org
Performance evaluation principles Performance is representated as a functional T depending on the receiver input distribution F General linear MU-MIMO channel Y A = H A S A + I A + N F = joint distribution of (H A, S A, I A, N) A = random set of active links MASCOT 2007 2
Role of the MU-MIMO scheduler The scheduler defines A depending in CQI feedback The scheduler shapes the statistics of H A and I A Given A, the PHY shapes the statistics of S A Evaluation of T(F ) must account for the dependencies among H A, S A, I A. MASCOT 2007 3
Three approaches considered Approach 1: Measure then simulate - Measure the MIMO channels with a channel sounder - Store the MIMO channels in a database - Run a Matlab simulation based on the stored channels Approach 2: Emulate using real frontends - Take two PCs running Matlab - Transmitter PC - Use Matlab to format a MIMO transmission block - Transmit it via a real MIMO transmitter frontend - Receiver PC - Receive it from a real MIMO receiver frontend - Implement the Detector/Decoder in Matlab - Evaluate performance using a side channel Approach 3: Real-time testbed MASCOT 2006 2
Approach 1: Measure then Simul. Advantages - Perfect repeatability of the performance evaluation - Very suitable for MU-MIMO performance evaluation - You can simulate with perfect channel state information (if you wish so) - Many environments can be handled - Great freedom in defining ensemble averages Disadvantages - The MIMO channel s behaviour must be - simulated in Matlab (convolution), and - idealised (analog Tx/Rx front-end behaviour!) - Channel measurement time, duration, and sampling rates do not match the simulated system - Non-realtime - It is questionable whether opportunistic MU-MIMO schemes based CSI feedback can be evaluated MASCOT 2006 3
Approach 2: Emulate Advantages: - Hic et nunc - MIMO channel is used rather than simulated - Real front-end behaviour is included in performance Disadvantages: - Performance evaluations are not perfectly repeatable - Great care must be taken when comparing two competing MU-MIMO schemes - MU-MIMO performance evaluation is significantly more costly than Single User case. - Non-realtime - It is questionable whether opportunistic schemes with CSI feedback can be evaluated MASCOT 2006 4
Approach 3: Real-time Testbed Advantages - Front-end effects included - Real MIMO channel behaviour - Opportunistic schemes with limited CSI feedback can be evaluated Disadvantages - Costly - Little flexibility - Performance evaluation not perfectly repeatable - It is difficult to emulate perfect CSI. - It is difficult to analyse a single effect MASCOT 2006 5
Example for Approach 1 Measure then simulate MEDAV s RUSK ATM channel sounder - 15 Tx elements (uniform circular array) - 8 Rx elements (uniform linear array) - Measurement band: 1940-2060 MHz - Urban, Sub-urban, Rich-scattering environments - http://www.ftw.at/measurements MASCOT 2006 6
Mobile Tx, Static Rx MASCOT 2006 7
Sub-urban environment Receiver s view @ Weikendorf
Outdoor Channel impulse response MASCOT 2006 9
Indoor rich scattering MASCOT 2006 10
Spatial re-sampling (1) MASCOT 2006 11
Spatial re-sampling (2) MASCOT 2006 12
Optimum receiver vs. Mismatched ML [9] G. Taricco and E. Biglieri: Space-time decoding with imperfect channel estimation, IEEE Trans. Wireless Commun. 4(4):1874-1888, July 2005 MASCOT 2006 13
Urban environment MASCOT 2006 14
Rich scattering environment MASCOT 2006 15
Example Approach 2: Emulate with real front-ends 4 Tx Front-ends 4 Rx Front-ends 2.6 GHz band, 20 MHz bandwidth MASCOT 2006 16
Experiment Set Up 3/42 Sebastian Caban scaban@nt.tuwien.ac.at Transmitter Roof or even higher
Experiment Set Up 5/42 Sebastian Caban scaban@nt.tuwien.ac.at Transmitter Roof or even higher Patch Antennas 20 dbm/antenna 2.5GHz 20dBm = 0.1 watt
Experiment Set Up 7/42 Sebastian Caban scaban@nt.tuwien.ac.at Channel real channel, urban environment TX 7 am
Experiment Set Up 8/42 Sebastian Caban scaban@nt.tuwien.ac.at Receiver Indoor
Experiment Set Up 9/42 Sebastian Caban scaban@nt.tuwien.ac.at Receiver Indoor 4 Monopoles
First Measurement 11/42 Sebastian Caban scaban@nt.tuwien.ac.at 13 E c /I or values 2 and 4 transmit antennas 2 CQI values 4, 6, 8, and 10 spreading-codes active 6552 realizations 4 hours net measurement time 600 GB of data received 4 days of work for 23 PCs
Measured Impulse Responses 12/42 Sebastian Caban scaban@nt.tuwien.ac.at TX1 RX1 TX2 RX1 approx 160 m TX1 RX2 TX2 RX2
Measured Impulse Responses 13/42 Sebastian Caban scaban@nt.tuwien.ac.at
BLER Results 36/42 Sebastian Caban scaban@nt.tuwien.ac.at Parameters: 2x2 system Code rate of 0.7 1000 realizations 10 Codes 45 equalizer taps 15 channel taps 95% confidence intervals
BLER Results 37/42 Sebastian Caban scaban@nt.tuwien.ac.at Parameters: 2x2 system Code rate of 0.7 6552 realizations 10 Codes 45 equalizer taps 15 channel taps 95% confidence intervals
BLER Results 38/42 Sebastian Caban scaban@nt.tuwien.ac.at Parameters: 2x[4,3,2] system Code rate of 0.7 6552 realizations 10 Codes 45 equalizer taps 15 channel taps
BLER Results 39/42 Sebastian Caban scaban@nt.tuwien.ac.at Parameters: [4x4, 2x2] system Code rate of 0.7 [1788,6552] realizations 10 Codes 45 equalizer taps 15 channel taps equal E B Question: power normalization in measurements?
BLER Results 40/42 Sebastian Caban scaban@nt.tuwien.ac.at Parameters: 2x2 system Code rate of 0.7 6552 realizations [4,6,8,10] Codes 45 equalizer taps 15 channel taps
Conclusions 41/42 Sebastian Caban scaban@nt.tuwien.ac.at If the Matlab-code works, one extended week 1 needed to measure the influence of a certain parameter We 1 = 7 days, 16 hours a day Mobilkom One