Full-duplex Wireless: From Experiments to Theory Achaleshwar Sahai, Melissa Duarte #, Evan Everett, Jingwen Bai, Gaurav Patel, Chris Dick* and Ashu Sabharwal Department of ECE Rice University # Now at EPFL/UCLA *Xilinx, Inc.
Half-duplex for Bi-directional Cellular FDD, TDD WiFi TDD Uplink and downlink on different bands or time-slots Difficult to transmit & receive in the same band at the same time.
Recent Full-duplex News Node 1 x 1 DAC BB Tx RF Radio c 1 DAC BB Tx Radio RF h z h ab y 1 ADC BB Rx Radio RF + Antenn Rice (Duarte et. al, 2010) Rice, AT&T (Duarte et. al, 2012) Stanford 1 (Choi et. al, 2010) l l phase shifter splitter Stanford II (Jain et. al, 2011) Princeton, NEC (Aryafar et. al, 2012)
Today s Story Line 2009-11 2011-12 2012-
Today s Story Line Experimental Full-duplex Explain the performance 2009-11 (Characterize bottlenecks) 2011-12 Distributed Full-duplex 2012-
Five 5MD* Sessions 2013 Distributed Full-duplex Passive Cancellation Bottleneck 2012 Q&A History Analog Canceller Bottleneck Main Challenge Main Techniques 2011 2010 *5MD = 5 Minutes & Done
Main Challenge of Full-duplex Node 1 Node 2 DAC RF UP RF UP DAC ADC RF Down RF Down ADC Self-interference is huge since h I >> h S 50-110dB larger than signal of interest, depending on inter-node distance
Bottleneck in Implementation Node 1 Node 2 DAC RF UP RF UP DAC ADC RF Down RF Down ADC Analog-to-Digital Conversion is the bottleneck
Analog to Digital Conversion RF Down ADC
Analog to Digital Conversion 1 8 6 4 2 Quantization of sum signal 0.2.4.6.8 1
1 8 6 4 2 0.2.4.6.8 1 Digital Cancellation Alone Insufficient: Signal of Interest Swamped Quantized sum signal 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 Signal of interest Only a few bits for the signal of interest Most baseband receivers need at least 6 bits For 12 bit ADC, INR(dB) > SNR(dB) + 35 db implies selfinterference is too strong
1 8 6 4 2 0.2.4.6.8 1 Digital Cancellation Alone Insufficient: Signal of Interest Swamped Quantized sum signal 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 Signal of interest Need to reduce strength of self-interference before ADC
Use Better ADCs Make ADC more capable? No Moore s law for ADC Limited by thermal noise No consensus on limits of current technology But consensus that progress is very slow Fundamentally limited by Heisenberg s uncertainty Walden 1999 Krone and Fettweis, 2009
Five 5MD Sessions 2013 Distributed Full-duplex Passive Cancellation Bottleneck 2012 Q&A History Analog Canceller Bottleneck Main Challenge Main Techniques 2011 2010
Recent Full-duplex Designs Node 1 x 1 DAC BB Tx RF Radio c 1 DAC BB Tx Radio RF h z h ab y 1 ADC BB Rx Radio RF + Antenn Rice (Duarte et. al, 2010) Rice, AT&T (Duarte et. al, 2012) Stanford 1 (Choi et. al, 2010) l l phase shifter splitter Stanford II (Jain et. al, 2011) Princeton, NEC (Aryafar et. al, 2012)
Main Principle of All Techniques Engineer Self-interference Channel
Self-Interference Channel Tx h I x[n] h S Rx baseband x[n] RF Up Tx Antenna h I Rx Antenna
Engineering h I : Passive Suppression x[n] RF Up Tx Antenna h I Rx Antenna Design antennas to increase propagation loss of h I
Engineering h I : Active Cancellation Self-interference signal is known Exploit it to cancel interference Many methods since 1998
Engineering h I : Active Analog Cancellation x[n] RF Up h I Cancellation Path + 0 Objective is to achieve exact 0 at the Rx antenna Cancellation path = negative of interfering path These techniques need analog parts
Engineering h I : Active Digital Cancellation x[n] RF Up h I + RF Down Baseband Cancellation Path Cancel interference at baseband Conceptually simpler requires no new parts Useless if interference is too strong (ADC bottleneck)
Rice WARP (warp.rice.edu) V1 V2 Up to 4-antenna MIMO per node Open-source design flows Operational networks out-of-the-box 125+ worldwide users, 140+ papers V3 (Mango)
Rice MIMO Full-duplex Median 85dB of total self-interference suppression [Duarte et. al, Dissertation and Journal 2012, arxiv 1210.1639] 20 MHz OFDM implementation Antennas on a laptop Analog cancellation Digital cancellation R1/ T3 T2 T1/ R2
Rice MIMO Full-duplex Median 85dB of total self-interference suppression [Duarte et. al, Dissertation and Journal 2012, arxiv 1210.1639] 20 MHz OFDM implementation Antennas on a laptop Analog cancellation Digital cancellation R1/ T3 T2 T1/ R2 5-node testbed with line-of-sight, 1, 2 and 3 walls links 5 PHY: 1x1, 2x1 Full-duplex, 2x1, 3x1 and 2x2 Half-duplex Month-long trials in regular office environments
Rate Performance 2x1 Full-duplex Rate > 2x2 Half-duplex for [15-30] db WiFi designed for [0,30] db 5 4 Ergodic Rate (bps/hz) 3 2 1 0 FD1X1 Experiments FD2X1 Experiments HD2X1 Experiments HD3X1 Experiments HD2X2 Experiments FD1X1 Linear Fit FD2X1 Linear Fit HD2X1 Linear Fit HD3X1 Linear Fit HD2X2 Linear Fit 0 10 20 30 40 50 60 SNR (db)
I am Convinced Full-duplex not just a headline Additional reduction in self-interference means higher rates How do we reduce self-interference?
Five 5MD Sessions 2013 Distributed Full-duplex Passive Cancellation Bottleneck 2012 Q&A History Analog Canceller Bottleneck Main Challenge Main Techniques 2011 2010
Node 1 x 1 All Analog Cancellers Are Imperfect. Question 1: Why? Why these numbers? DAC BB Tx RF Radio c 1 DAC BB Tx Radio RF h z h ab y 1 ADC BB Rx Radio RF + Antenn 34dB 30dB l l phase shifter splitter 45 db 45 db
Total Cancellation < Sum of Max Question 2: Why? x[n] h I Analog Canceller ADC Digital Canceller Digital cancellation by itself achieves ~30dB Analog is 30-45dB But total is closer to 35-50dB, not 60-75dB More analog cancels digital cancels less See Sahai, Patel, Dick and Sabharwal, 2012 (arxiv 1212.5462)
Analog Cancellers Bottleneck Phase Noise
Engineering h I : Analog Cancellation x[n] RF Up h I Cancellation Path + 0 Objective is to achieve exact 0 at the Rx antenna Cancellation path = negative of interfering path These techniques need analog parts
At-antenna Analog Cancellation RF Up h I x[n] + f( ) RF Up Pre-mixer x[n] RF Up h I + Post-mixer RF Up g( )
Classifying Recent Designs Node 1 x 1 DAC BB Tx RF Radio c 1 DAC BB Tx Radio RF h z h ab y 1 ADC BB Rx Radio RF + Antenn Pre-mixer Post-mixer l l phase shifter splitter Post-mixer Post-mixer
Question 1: Why Rice Analog Cancels 34dB? Errors in up-conversion RF Up h I x[n] + f( ) RF Up Residual Errors in implementing ideal f( )
Single-tap Delay Channel RF Up x[n] + f( ) RF Up h (t )
Errors in Up-conversion: Phase Noise x[n] + j(!t+ (t)) e x h (t ) f( ) x All oscillators are imperfect, Random phase variations
Perfect Channel Estimate x[n] j(!t+ (t)) + e x f( ) x h (t ) residual f( ) = h (t )e j! Same phase noise goes through different channels residual = hx(t )e j!(t ) e j (t ) e j (t) jhx(t )e j!(t ) ( (t ) (t)) time-varying
Perfect Channel Estimate x[n] j(!t+ (t)) + e x f( ) x h (t ) residual f( ) = h (t )e j! Same phase noise goes through different channels residual = hx(t )e j!(t ) e j (t ) e j (t) jhx(t )e j!(t ) ( (t ) (t)) E residual 2 2 h 2 2 (1 R ( )) + 2 noise
No Channel Estimation Error 70 65 Cancelation (in db) 60 55 50 45 40 35 SigGen! 1 10 log 10 2 2 WARP 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 (in degrees)
Measured Versus Predicted 70 Cancelation (in db) 65 60 55 50 45 40 35 X SigGen Pre-mixer: 56dB ( perfect )! 1 10 log 10 2 2 WARP 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 (in degrees) X Pre-mixer: 36dB ( perfect ) Experiments over wire ( near-perfect channel estimates)
Impact of Channel Estimation Error 70 Cancelation (in db) 65 60 55 50 45 40 35 X X SigGen Pre-radio: 56dB ( perfect ) Duarte 10 : 50dB (imperfect)! 1 10 log 10 2 2 WARP 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 (in degrees) X X Pre-radio: 36dB ( perfect ) Duarte 10: 34dB (imperfect)
Possible to Circumvent Phase Noise? x h (t ) x[n] + j(!t+ (t)) e x g( ) g( ) = h (t ) In theory, possible to achieve zero residual.
Most post-radio report 30-45dB cancellation (not infinite db) Node 1 x 1 DAC BB Tx RF Radio c 1 DAC BB Tx Radio RF h z h ab y 1 ADC BB Rx Radio RF + Antenn Pre-mixer 34dB Post-mixer 30dB l l phase shifter splitter Post-mixer 45 db Post-mixer 45 db
Recall the Numbers Node 1 x 1 DAC BB Tx RF Radio c 1 DAC BB Tx Radio RF h z h ab y 1 ADC BB Rx Radio RF + Antenn 34dB Pre-mixer Post-mixer l l phase shifter splitter 45 db Post-mixer WARP Post-mixer 45 db
Pre Versus Post 70 Cancelation (in db) 65 60 55 50 45 40 35 X X SigGen Pre-radio: 56dB ( perfect ) Duarte 10 : 50dB (imperfect)! 1 10 log 10 2 2 WARP 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 (in degrees) X X X Post-radio: 50dB (narrowband) Stanford 11: 45dB (wideband with flat fading) Not measured by us Pre-radio: 36dB ( perfect ) Duarte 10: 34dB (imperfect) Sahai, Patel, Dick, Sabharwal, 2012 (arxiv 1212.5462)
Five 5MD Sessions 2013 Distributed Full-duplex Passive Cancellation Bottleneck 2012 Q&A History Analog Canceller Bottleneck Main Challenge Main Techniques 2011 2010
Rice MIMO Full-duplex Median 85dB of total self-interference suppression [Duarte et. al, Dissertation and Journal 2012, arxiv 1210.1639] 20 MHz OFDM implementation Antennas on a laptop Analog cancellation Digital cancellation R1/ T3 T2 T1/ R2 5-node testbed with line-of-sight, 1, 2 and 3 walls links 5 PHY: 1x1, 2x1 Full-duplex, 2x1, 3x1 and 2x2 Half-duplex Month-long trials in regular office environments
Rice MIMO Full-duplex Median 85dB of total self-interference suppression T2 65 db Passive 20 db Active R1/ T3 T1/ R2 Passive dominates the total cancellation!
Self-Interference Channel Local scattering Tx Direct path Rx Full-duplex Node
Self-Interference Channel Local scattering Tx Direct path Rx Full-duplex Node Can actively address this
Option 1: Antenna Placement Tx Rx Tx d Rx Possibility for mobile nodes (Duarte, et. al. 2012)
Option 1I: RF Absorbers Tx Rx Full-duplex Node Possibility for infrastructure nodes
Option 1II: Cross Polarization H V Tx Rx Full-duplex Node Possibility for mobile & infrastructure nodes
Option 1V: Directional Communication Tx Rx Full-duplex Node Possibility for infrastructure nodes
Infrastructure Nodes B AP RF Absorber Dual-polarized directional antennas
Tests 50 cm 50 cm 35 cm 50 cm 50 cm 50 cm 50 cm 35 cm 35 ion I: 90 beamwidth (b) Configuration II: 90 beamwidth (c) Configuration III: 90 beamwidth Configuration I: 90 beamwidth (b) Configuration II: 90 beamwidth (c)separation, Configuration 90(a)beam separation, antennas, 60 beam separation, antennas,(c) 60 Configuration beam figuration I: 90 beamwidth (b) Configuration II: 90 beamwidth III: 90 III be antennas, beam separation, antennas, beam separation, antennas, 60sepa bea enna 50 cm antenna separation. 35 cm antenna separation. nas, separation. 90 beam90 separation, antennas, 60 beam60separation, antennas, 60 beam 50 cmseparation. antenna separation. 50 cmseparation. antenna separation. 35 cm antenna se m antenna 50 cm antenna 35 cm antenna separation D : Directional DA : Directional + Absorber DC : Directional + Cross-pol DCA : Directional + Cross-pol + Absorber 50 cm (d) Configuration IV: 50 cm 50 cm Omnidirectional antennas, Configuration IV: 50 (d) cm Configuration antenna(d)separation. IV: 35 cm (e) Configuration V: 35 cm Omnidirectional antennas,35 cm (e) Configuration V: 35 cm antenna separation. (e) Configuration V: Omnidirectional antennas, Omnidirectional antennas, Omnidirectional antennas, Omnidirectional antennas, 50 cm antenna separation. 35 cm antenna separation. cmthe antenna separation. 35 cmwe antenna separation. ures evaluated50 in passive suppression characterization. do not designate one antennas
Reflections are the Bottleneck 80 NASA Anechoic Chamber 74 db 70 A Very Reflective Room Room Passive Suppression (db) 60 40 Passive Suppression (db) 60 50 40 30 46 db 20 Config I Config II Config III Reflections harder to control for mobile Easier for infrastructure nodes Stationary Carefully placed 20 Config I Config II Config III
We Tried This Up to 120 meters
Self-Interference Close to Noise Floor 1 0.75 w/o Cross pol: passive w/o Cross pol: passive + active w/ Cross pol: passive w/ Cross pol: passive + active DCA + Active Mean 95dB+ CDF 0.5 0.25 0 50 60 70 80 90 100 Self Interference Supression (db) Everett, Sahai, Sabharwal, 2013 (arxiv 1302.2185)
Five 5MD Sessions 2013 Distributed Full-duplex Passive Cancellation Bottleneck 2012 Q&A History Analog Canceller Bottleneck Main Challenge Main Techniques 2011 2010
First Potential Use Case Full-duplex Infrastructure I Half-duplex M 1 M 2 Inter-node Interference Half-duplex
New Bad News Inter-node Interference Very weak Very strong Multiplexing gain New bad news! Dead Zone Strength of Inter-node Interference
New Question How do we achieve multiplexing gain of 2? (Get the full-duplex benefit)
Compare These Two Scenarios I I M M 1 M 2 Tx and Rx are co-located Can we mimic it here?
Mobile Devices Today Cellular Band: 800, 850, 900, 1800, 1900, 2100 MHz WiFi : 2.4GHz and 5GHz iphone 5
Add Side Channel: ISM-in-Cellular I Cellular M 1 M 2 ISM
Near-optimal Scheme: Bin-and-cancel I Cellular M 1 M 2 ISM Use ISM side-channel to cancel the main channel interference Finite bit optimal for all SNR regimes Bai and Sabharwal, arxiv 1212.5300, December 2012
Parameter of Interest I Cellular M 1 M 2 ISM Bandwidth Ratio B = BW side BW main
B = 0.5 Red: Bin-and-cancel Multiplexing gain Blue: Decodeand-cancel 50% Cyan: Best achievable scheme without side channel log INR SNR
B ~ 5 Red: Bin-and-cancel Multiplexing gain Blue: Decodeand-cancel Cyan: Best achievable scheme without side channel log INR SNR Bai, Sabharwal, 2012 (arxiv 1212.5300)
So What Next? (argos.rice.edu) Lots of interesting ideas Both on infrastructure and mobile side -+./01*% -'./0'**+0!"#$""%&!"#$% &'()*+, 5+6#)& 7.*1+$#..1$*) 24.$ '()*+(,-*(#.!"#$%& '() 64-element Argos!"#$%& '()*+(,-*(#. /*01+.1*& 23(*$0
Questions & Comments? JSAC Special Issue on Full-duplex Wireless, Oct 1 2013 WARP Project - http://warp.rice.edu ARGOS Project http://argos.rice.edu