Experiment-Driven Characterization of Full-Duplex Wireless Systems

Experiment-Driven Characterization of Full-Duplex Wireless Systems Melissa Duarte Advisor: Ashutosh Sabhawal Department of ECE Rice University August 04 2011 1

Full-Duplex Wireless Node 1 Node 2 Same time same frequency band Assumed to be impossible due to large self interference Revisit this assumption using techniques for interference cancellation Can full-duplex achieve higher data rates than half-duplex? Characterize amount of cancellation and achievable rate performance 2

Interference Problem x h I Node 1 Node 2 Theory Interference >> Signal Strong interference regime Interference is known, estimate channel, cancel, done 3

Interference Problem x h I Node 1 Node 2 Implementation 4

Interference Problem Node 1 x(t)e jwt x (t)e jwt x[n] DAC Tx RF h I h S y[n] ADC y(t) Rx RF y(t)e jwt Node 2 Implementation 5

Interference Problem Node 1 x(t)e jwt x (t)e jwt x[n] DAC Tx RF h I h S y[n] ADC y(t) Rx RF y(t)e jwt Node 2 Received signal Implementation y(t) =h S x (t)+h I x(t)+z(t) Signal of interest Interfering signal 6

Interference Problem Node 1 x(t)e jwt x (t)e jwt x[n] DAC Tx RF h I h S y[n] ADC y(t) Rx RF y(t)e jwt Node 2 Implementation Received signal y(t) =h S x (t)+h I x(t)+z(t) Quantized received signal y[n] =h S x [n]+h I x[n]+z[n] Signal of interest Interfering signal 7

Interference Problem Node 1 x(t)e jwt x (t)e jwt x[n] DAC Tx RF h I h S y[n] ADC y(t) Rx RF y(t)e jwt Node 2 Implementation Received signal Quantized received signal After removing interference y(t) =h S x (t)+h I x(t)+z(t) y[n] =h S x [n]+h I x[n]+z[n] y[n] =h S x [n]+z[n] Signal of interest Interfering signal Quantization noise 8

Interference Problem Implementation Interference >> Signal : Quantization noise Full-duplex assumed to be impossible due to large self interference Real systems SIR -100 db -20 db Distance between nodes decreases Outdoor Indoor Cellular Bluetooth, WiFi More than 20 db cancellation has been reported in radar systems Revisit this assumption using passive and active techniques for interference cancellation 9

Passive Cancellation Antenna Separation Antenna Directionality. Antenna Cancellation Device Cancellation 10

Passive Cancellation Antenna Separation Separation d between same node Tx and Rx antennas d Antenna Directionality Antenna Cancellation Device Cancellation 11

Passive Cancellation Antenna Separation Separation d between same node Tx and Rx antennas d Antenna Directionality Used in Full-duplex Relays S R D Everett et. al. Empowering Full-Duplex Wireless Communication by Exploiting Directional Diversity. Asilomar 2011. Haneda et. al. Measurement of Loop-Back Interference Channels for Outdoor-to-Indoor Full-Duplex Radio Relays. Eucap 2010. Antenna Cancellation 2 Tx and 1Rx per node, Tx at d and d+λ/2 Choi et al. Achieving Single Channel, Full Duplex Wireless Communication. Mobicom 2010. 1 2 d+λ/2 d Device Cancellation 13

Passive Cancellation Antenna Separation Separation d between same node Tx and Rx antennas We use antenna separation with d = 10cm, 20cm, 40cm Worse case interference Minimum resources for passive cancellation Antenna Directionality Used in Full-duplex Relays Everett et. al. Empowering Full-Duplex Wireless Communication by Exploiting Directional Diversity. Asilomar 2011. Haneda et. al. Measurement of Loop-Back Interference Channels for Outdoor-to-Indoor Full-Duplex Radio Relays. Eucap 2010. S R D d Antenna Cancellation 2 Tx and 1Rx per node, Tx at d and d+λ/2 Choi et al. Achieving Single Channel, Full Duplex Wireless Communication. Mobicom 2010. 1 2 d+λ/2 d Device Cancellation Place antennas at opposite sides of the device Sahai et al. Pushing the Limits of Full-Duplex: Design and Real-Time Implementation. Tech. Report 2011. Rice University. 15

Active Analog Cancellation Using QHx220 chip Using extra Tx RF chain (without a power amplifier) 16

Active Analog Cancellation Using QHx220 chip Tuning algorithm to control gain and delay that chip applies to its input Suitable for wideband frequency flat h I x y DAC Tx RF control ADC Rx RF QHx 220 + h I Choi et al. Achieving Single Channel, Full Duplex Wireless Communication. Mobicom 2010. Radunovic et al. Rethinking Indoor Wireless: Low Power, Low Frequency, Full-Duplex. Tech. Report Microsoft Research 2009. Using extra Tx RF chain (without a power amplifier) 17

Active Analog Cancellation Using QHx220 chip Tuning algorithm to control gain and delay that chip applies to its input Suitable for wideband frequency flat h I x y DAC control ADC Tx RF DAC Rx RF QHx 220 + h I Choi et al. Achieving Single Channel, Full Duplex Wireless Communication. Mobicom 2010. Radunovic et al. Rethinking Indoor Wireless: Low Power, Low Frequency, Full-Duplex. Tech. Report Microsoft Research 2009. Using extra Tx RF chain (without a power amplifier) x DAC Tx RF Estimate h I and design c for analog cancellation Suitable for wideband frequency flat and frequency selective h I c y DAC ADC Tx RF Rx RF + h I Uses off-the-shelf MIMO radios Duarte et al. Full-Duplex Wireless Communications using Off-The-Shelf Radios: Feasibility and First Results. Asilomar 2010. Sahai et al. Pushing the Limits of Full-Duplex: Design and Real-Time Implementation. Tech. Report 2011. Rice University. 18

Active Analog Cancellation Using QHx220 chip Tuning algorithm to control gain and delay that chip applies to its input Suitable for wideband frequency flat h I x y DAC control ADC Tx RF DAC Rx RF QHx 220 + h I Choi et al. Achieving Single Channel, Full Duplex Wireless Communication. Mobicom 2010. Radunovic et al. Rethinking Indoor Wireless: Low Power, Low Frequency, Full-Duplex. Tech. Report Microsoft Research 2009. Used in our experiments Using extra Tx RF chain (without a power amplifier) x DAC Tx RF Estimate h I and design c for analog cancellation Suitable for wideband frequency flat and frequency selective h I c y DAC ADC Tx RF Rx RF + Uses off-the-shelf MIMO radios Duarte et al. Full-Duplex Wireless Communications using Off-The-Shelf Radios: Feasibility and First Results. Asilomar 2010. Sahai et al. Pushing the Limits of Full-Duplex: Design and Real-Time Implementation. Tech. Report 2011. Rice University. 19

Active Digital Cancellation Without analog cancellation Combined with analog cancellation

Active Digital Cancellation Without analog cancellation Estimate h I and cancel h I x in the digital domain x y DAC ADC Tx RF Rx RF h I Combined with analog cancellation 21

Active Digital Cancellation Without analog cancellation Estimate h I and cancel h I x in the digital domain x y DAC ADC Tx RF Rx RF h I Combined with analog cancellation x DAC Tx RF Estimate residual interference and cancel in the digital domain c y DAC ADC Tx RF Rx RF + h I We have considered the two options above Allows us to characterize the effect in total cancellation when concatenating cancellation mechanisms Duarte et al. Experiment-Driven Characterization of Full-duplex Wireless Systems. Submitted to IEEE Trans. Wireless 2011. 23

Full-Duplex Systems Considered We have implemented the following full-duplex systems Antenna Separation + Digital Cancellation Antenna Separation + Analog Cancellation Antenna Separation + Analog Cancellation + Digital Cancellation 24

Summary of Results Characterization of self-interference cancellation mechanisms Amount of cancellation Impact on full-duplex achievable rate performance Comparison with half-duplex systems Demonstrated that full-duplex can achieve higher rates than half-duplex Duarte et al. Full-Duplex Wireless Communications using Off-The-Shelf Radios: Feasibility and First Results. Asilomar 2010. Duarte et al. Experiment-Driven Characterization of Full-duplex Wireless Systems. Submitted to IEEE Trans. Wireless 2011. 25

Experiment Setup Node 1 Node 2 8.5 m 8.5 m Node 2 Node 1 26

Experiment Setup d d = 10 cm, 20 cm, 40 cm d Node 1 Node 2 8.5 m 20 cm 8.5 m Node 2 Node 1 27

Experiment Setup P = [0,5,10,15] dbm d d = 10 cm, 20 cm, 40 cm d Node 1 Node 2 8.5 m 20 cm 8.5 m Node 2 Node 1 28

Experiment Setup Node 1 x c DAC DAC P = [0,5,10,15] dbm Tx RF d d = 10 cm, 20 cm, 40 cm d Node Tx RF1 Node 2 y ADC Rx RF + 8.5 m WARP with 3 radios 20 cm WARPLab = WARP + Matlab, to generate/analyze signals Node 1 8.5 m Node 2 Narrowband tests, 0.65 MHz Recent extension to OFDM 10MHz @ Rice Sahai et al. Pushing the Limits of Full-Duplex: Design and Real-Time Implementation. Tech. Report 2011. Rice University. 29

Experiment Setup Node 1 x c DAC DAC P = [0,5,10,15] dbm Tx RF d d = 10 cm, 20 cm, 40 cm d Node Tx RF1 Node 2 y ADC Rx RF + RF combiner 8.5 m WARP with 3 radios 20 cm WARPLab = WARP + Matlab, to generate/analyze signals Node 1 8.5 m Node 2 Narrowband tests, 0.65 MHz Recent extension to OFDM 10MHz @ Rice Sahai et al. Pushing the Limits of Full-Duplex: Design and Real-Time Implementation. Tech. Report 2011. Rice University. RF combiner 30

Characterization of Average Cancellation Digital cancellation Analog cancellation Analog and digital cancellation 31

Characterization of Average Cancellation Digital cancellation Analog cancellation Analog and digital cancellation x DAC Tx RF y ADC Rx RF P I 32

Characterization of Average Cancellation Digital cancellation Analog cancellation Analog and digital cancellation x DAC Tx RF c DAC Tx RF y ADC Rx RF + P I P IAC 33

Characterization of Average Cancellation Digital cancellation Analog cancellation α AC =P I P IAC Analog and digital cancellation x DAC Tx RF c DAC Tx RF y ADC Rx RF + P I P IAC 34

Characterization of Average Cancellation Digital cancellation α DC =P I P IDC Analog cancellation α AC =P I P IAC Analog and digital cancellation x DAC Tx RF y DC DC ADC Rx RF P I P IDC 35

Characterization of Average Cancellation Digital cancellation α DC =P I P IDC Analog cancellation α AC =P I P IAC Analog and digital cancellation α ACDC =P I P IACDC x DAC Tx RF c DAC Tx RF y ACDC DC ADC Rx RF + P I P IACDC 36

Characterization of Average Cancellation Digital cancellation Analog cancellation Analog and digital cancellation α DC =P I P IDC α AC =P I P IAC α ACDC =P I P IACDC!#!#!# $" $" $",-01203443, $# %",-01203443, $# %",-01203443, $# %" %# %# %#!!"!!#!$"!$#!%"!%# & '() +,-./!!"!!#!$"!$#!%"!%# & '() +,-./!!"!!#!$"!$#!%"!%# & '() +,-./ 37

Characterization of Average Cancellation,-01203443, Digital cancellation α DC =P I P IDC!#! 37839:.32;< 56 $"! 56 0=2<;12;.=,34! 4:2319.=,34 56 $# %" %#!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443, Analog cancellation α AC =P I P IAC!# $" $# %"! 37839:.32;< 56 %#! 0=2<;12;.=,34 56! 56 4:2319.=,34!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443, Analog and digital cancellation α ACDC =P I P IACDC!# $" $# %"! 3893:;.32<= 5676 %#! 0>2=<12<.>,34 5676! 5676 4;231:.>,34!!"!!#!$"!$#!%"!%# & '() +,-./ We want a simple model for the average cancellation Option 1: Fit the data to a constant model Option 2: Fit the data to a linear model 40

Characterization of Average Cancellation,-01203443, Digital cancellation α DC =P I P IDC!#! 37839:.32;< 56 $"! 56 0=2<;12;.=,34! 4:2319.=,34 56 $# %" %#!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443, Analog cancellation α AC =P I P IAC!# $" $# %"! 37839:.32;< 56 %#! 0=2<;12;.=,34 56! 56 4:2319.=,34!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443, Analog and digital cancellation α ACDC =P I P IACDC!# $" $# %"! 3893:;.32<= 5676 %#! 0>2=<12<.>,34 5676! 5676 4;231:.>,34!!"!!#!$"!$#!%"!%# & '() +,-./ Result 1: As received interference power P I increases amount of cancellation increases Reason: Cancellation is based on channel measurement Higher P I means higher SNR for channel estimation Better estimation and hence increased cancellation 41

Characterization of Average Cancellation,-01203443, Digital cancellation α DC =P I P IDC!#! 37839:.32;< 56 $"! 56 0=2<;12;.=,34! 4:2319.=,34 56 $# %" %#!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443, Analog cancellation α AC =P I P IAC!# $" $# %"! 37839:.32;< 56 %#! 0=2<;12;.=,34 56! 56 4:2319.=,34!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443, Analog and digital cancellation α ACDC =P I P IACDC!# $" $# %"! 3893:;.32<= 5676 %#! 0>2=<12<.>,34 5676! 5676 4;231:.>,34!!"!!#!$"!$#!%"!%# & '() +,-./ Result 2: (a) Concatenation of cancellation mechanisms does not result in a sum of their individual cancellations 42

Characterization of Average Cancellation Digital cancellation Analog cancellation Analog and digital cancellation,-01203443, α DC =P I P IDC α AC =P I P IAC α ACDC =P I P IACDC!#! 37839:.32;< 56 $"! 56 0=2<;12;.=,34! 4:2319.=,34 56 $# %" %#!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443,!# $" $# %"! 37839:.32;< 56 %#! 0=2<;12;.=,34 56! 56 4:2319.=,34!!"!!#!$"!$#!%"!%# & '() +,-./,-01203443,!# $" $# %"! 3893:;.32<= 5676 %#! 0>2=<12<.>,34 5676! 5676 4;231:.>,34!!"!!#!$"!$#!%"!%# & '() +,-./ Max cancellation : 27 db Max cancellation : 35 db Max cancellation : 36 db Result 2: (a) Concatenation of cancellation mechanisms does not result in a sum of their individual cancellations 43

Characterization of Average Cancellation Average performance α DC = α ACDC α AC Per frame performance α DC [f] =α ACDC [f] α AC [f]! 2+, -./01 αdc (" ' & %! "! 2+, -34536,7389: ;<8! 2+, =,8! 2+,!! -!"!# $" $#! )+, -./01 Result 2: α AC - Probability(αDC[f] > 0) (%) Probability (! DC,i [f] > 0) (%) 100 90 80 70 60 50 40 30 20 10 0 15 20 25 30 35 40! AC,i [f] (db) α AC [f] (a) Concatenation of cancellation mechanisms does not result in a sum of their individual cancellations (b) As the performance of analog cancellation gets better, the effectiveness of digital cancellation after analog cancellation reduces (observed on average and per frame) 45

Characterization of Average Cancellation Average performance α DC = α ACDC α AC Per frame performance α DC [f] =α ACDC [f] α AC [f]! 2+, -./01 αdc (" ' & %! "!! -!"!# $" $#! )+, -./01 α AC Reasons for Result 2:! 2+, -34536,7389: ;<8! 2+, =,8! 2+, - Probability(αDC[f] > 0) (%) Probability (! DC,i [f] > 0) (%) 100 90 80 70 60 50 40 30 20 10 0 15 20 25 30 35 40! AC,i [f] (db) α AC [f] As residual interference becomes smaller the effectiveness of cancelling the residual reduces If analog cancellation could achieve db of cancellation then digital cancellation would be unnecessary Furthermore applying digital cancellation would increase the noise 46

Computation of Achievable Sum Rate Achievable Sum Rate (ASR) b/s/hz computed based on post processing SINR Q s Compute SINR per frame SINR[f] = E [ s 2 ] E [ s ŝ 2 ] ŝ I Compute achievable rate AR = E [log(1 + SINR[f])] Sum rate full-duplex ASR = AR 12 + AR 21 Sum rate half-duplex ASR = 1 2 AR 12 + 1 2 AR 21 48

Achievable Sum Rate Analysis x Node 1 Node 2 y P h I SIR = P S P I h S P x Result 3: For a fixed SIR at Rx antenna, increasing the transmit power increases the total achievable rate. Reasons for Result 3... 49

Achievable Sum Rate Analysis x P h I Node 1 Node 2 y SIR = P S P I Without active cancellation h S y = h S x + h I x + z P x With active cancellation y = h S x + ( ) h I ĥi x + z With active cancellation rewrite as h R y = h S x + h R Ω α x + z : normalized residual channel Ω : due to antenna separation 50 α SINR = : due to active cancellation Duarte et al. Experiment-Driven Characterization of Full-duplex Wireless Systems. Submitted to IEEE Trans. Wireless 2011. 1 1 αsir + 1 SNR

Achievable Sum Rate Analysis x Node 1 Node 2 y P h I SIR = P S P I h S P x Achievable sum rate ASR = log (1 + SINR 1 ) + log (1 + SINR 2 ) SINR i = 1 1 α i SIR i + 1 SNR i 51

Achievable Sum Rate Analysis x Node 1 Node 2 y Achievable sum rate P h I SIR = P S P I h S P ASR = log (1 + SINR 1 ) + log (1 + SINR 2 ) SINR i = Result 3: For a fixed SIR at Rx antenna, increasing the transmit power increases the total achievable rate. x 1 1 α i SIR i + 1 SNR i Reasons for Result 3: If P at both nodes increases by same amount then SIR doesn t change α increases (from Result 1) SINR increases SNR increases 52 Achievable rate increases

Achievable Sum Rate Analysis Result 3 demonstrated in experiments and simulation 3456789:;7-<=1->9?7-.:@A@BC2 #$ #" # #) #( ## #! ' & % /-D-!-41-7EF7G617H?A-IJ!3K /-D-(!-41-7EF7G617H?A-IJ!3K /-D-#!-41-7EF7G617H?A-IJ!3KJK /-D-(!-41-A61=;9?6LH-IJ!3K - FD-AC: Full-duplex with antenna separation and Analog Cancellation FD-ACDC: Full-duplex with antenna separation and combined Analog and Digital Cancellation $ - "! " #! #" + -./012, Simulation results obtained using model Setting y = h S x + h R Ω α x + z SNR = SIR = Measured in experiments α = Linear fit Linear fit model seems reasonably accurate 54 SINR = 1 1 αsir + 1 SNR

Result 4: Best performance is achieved when applying digital cancellation selectively based on measured suppression values Reasons for Result 4: Achievable Sum Rate Analysis For each frame decide if digital cancellation after analog cancellation should be applied or not as follows Use training at the beginning of the frame to estimate α AC [f] and Apply digital cancellation after analog cancellation during frame f only if α ACDC [f] α AC [f] > 0 α ACDC [f] 55

Result 4: Best performance is achieved when applying digital cancellation selectively based on measured suppression values Reasons for Result 4: Follows from Result 2: (a) Achievable Sum Rate Analysis Concatenation of cancellation mechanisms does not result in a sum of their individual cancellations (b) As the performance of analog cancellation gets better, the effectiveness of digital cancellation after analog cancellation reduces (observed on average and per frame) For each frame decide if digital cancellation after analog cancellation should be applied or not as follows Use training at the beginning of the frame to estimate α AC [f] and α ACDC [f] Apply digital cancellation after analog cancellation during frame f only if α ACDC [f] α AC [f] > 0 56

Achievable Sum Rate Analysis Result 4: Best performance is achieved when applying digital cancellation selectively based on measured suppression values Experiment results 0123456784"9:.";6<4"+7=>=?@/ $( $' $& $% $! ( ' & % AB!0C AB!0CBC AB!0C9BC " FD-AC: Full-duplex with antenna separation and Analog Cancellation FD-ACDC: Full-duplex with antenna separation and combined Analog and Digital Cancellation FD-ACDC: Full-duplex with antenna separation and combined Analog and Selective Digital Cancellation! "!" # $! $# ) "+,-./ Benefits of selective digital cancellation Uses digital cancellation as a safety net in frames where analog cancellation delivers poor performance Avoids adding noise to the system when analog cancellation delivers good performance Results in largest average achievable sum rate 58

Full-Duplex vs. Half-Duplex Compare half-duplex and full-duplex systems that use same resources Full-duplex with analog cancellation Half-duplex 2x1 Alamouti Antennas per node 2 2 Tx RF radios per node 2 2 Rx RF radios per node 1 1 Tx power per antenna P P Node 1 Node 2 x DAC Tx RF Tx RF DAC x c DAC Tx RF Tx RF DAC c y ADC Rx RF + + Rx RF ADC y 62

Full-Duplex vs. Half-Duplex d = 10 cm d = 20 cm d = 40 cm #" ' -./0123451'67+'8391'(4:;:<=, #! " >755!)7?51@'1@?1A0+1B9; <35C!)7?51@'1@?1A0+1B9; >755!)7?51@';0+75390DB! '! " #! #" $ %& '()+, 63

Full-Duplex vs. Half-Duplex d = 10 cm d = 20 cm d = 40 cm #" ' #" ' -./0123451'67+'8391'(4:;:<=, #! " >755!)7?51@'1@?1A0+1B9; <35C!)7?51@'1@?1A0+1B9; >755!)7?51@';0+75390DB! '! " #! #" $ %& '()+, -./0123451'67+'8391'(4:;:<=, #! " >755!)7?51@'1@?1A0+1B9; <35C!)7?51@'1@?1A0+1B9; >755!)7?51@';0+75390DB! '! " #! #" $ %& '()+, Full-duplex can achieve higher rates than half-duplex The linear fit model is reasonable accurate 64

Full-Duplex vs. Half-Duplex d = 10 cm d = 20 cm d = 40 cm #" ' #" ' #" ' -./0123451'67+'8391'(4:;:<=, #! " >755!)7?51@'1@?1A0+1B9; <35C!)7?51@'1@?1A0+1B9; >755!)7?51@';0+75390DB! '! " #! #" $ %& '()+, -./0123451'67+'8391'(4:;:<=, #! " >755!)7?51@'1@?1A0+1B9; <35C!)7?51@'1@?1A0+1B9; >755!)7?51@';0+75390DB! '! " #! #" $ %& '()+, -./0123451'67+'8391'(4:;:<=, #! " >755!)7?51@'1@?1A0+1B9; <35C!)7?51@'1@?1A0+1B9; >755!)7?51@';0+75390DB! '! " #! #" $ %& '()+, Full-duplex can achieve higher rates than half-duplex The linear fit model is reasonable accurate Close antennas imply in some scenarios half-duplex wins-out 65

Conclusions Full-duplex can achieve higher rates than half-duplex. Amount of active cancellation increases as the received self-interference power increases. At a constant SIR@Rx antenna more interference is actually good. It allows better cancellation and thus improved rates. Digital cancellation is more effective when applied selectively after analog cancellation. Recent and ongoing work at Rice Asynchronous full-duplex. Receive-while-sending. (not send-while-receiving) Antenna design and MAC protocols. MIMO and OFDM analysis. Everett et. al. Empowering Full-Duplex Wireless Communication by Exploiting Directional Diversity. Asilomar 2011. Sahai et al. Pushing the Limits of Full-Duplex: Design and Real-Time Implementation. Tech. Report 2011. Rice University. 68