Full Duplex Wireless

Philip LevisStanford University

@Princeton 10.12.2012

also Mayank Jain, Jung Il Choi, Sachin Katti, Kannan Srinivasan, Taemin Kim, Dinesh Bharadia, Prasun Sinha, and Siddarth Seth.

1

2

Chuck Thacker InterviewCommunications of the ACM, July 2010

3

“It is generally not possible for radios to receive and transmit on the same frequency band because of the interference that results. Thus, bidirectional systems must separate the uplink and downlink channels into orthogonal signaling dimensions, typically using time or frequency dimensions.” - Andrea Goldsmith, “Wireless Communications,” Cambridge Press, 2005.

• We’ve built working full duplex radios‣ Breaks a fundamental assumption in wireless‣ Current limitations: transmit power: ≤ WiFi

(200mW), bandwidth (≤ 20MHz)

• How full duplex could transform wireless

• Open technical barriers and challenges

Talk In a Nutshell

Why Half Duplex?

TX RX TX RX

• Self-interference is millions to billions (60-90dB) stronger than received signal

• If only we could “subtract” our transmission...

• Digital cancellation: subtracting known interference digital samples from received digital samples.

ZigZag[1], Analog Network Coding[2] etc.

• Hardware cancellation: RF noise cancellation circuits with transmit signal as noise reference

Radunovic et al.[3]

Existing Techniques

6

[1] Gollakota et al. “ZigZag Decoding: Combating Hidden Terminals in Wireless Networks”, ACM SIGCOMM 2008[2] Katti et al. “Embracing Wireless Interference: Analog Network Coding”, ACM SIGCOMM 2007[3] Radunovic et al. , "Rethinking Indoor Wireless: Lower Power, Low Frequency, Full-duplex", WiMesh (SECON Workshop),, 2010

• Digital cancellation: subtracting known interference digital samples from received digital samples.

ZigZag[1], Analog Network Coding[2] etc.

Ineffective if ADC saturates

• Hardware cancellation: RF noise cancellation circuits with transmit signal as noise reference

Radunovic et al.[3]

Existing Techniques

7

[1] Gollakota et al. “ZigZag Decoding: Combating Hidden Terminals in Wireless Networks”, ACM SIGCOMM 2008[2] Katti et al. “Embracing Wireless Interference: Analog Network Coding”, ACM SIGCOMM 2007[3] Radunovic et al. , "Rethinking Indoor Wireless: Lower Power, Low Frequency, Full-duplex", WiMesh (SECON Workshop),, 2010

These are not enough: 25dB +15dB = 40dB

• Digital cancellation: subtracting known interference digital samples from received digital samples.

ZigZag[1], Analog Network Coding[2] etc. ~15dB

Ineffective if ADC saturates

• Hardware cancellation: RF noise cancellation circuits with transmit signal as noise reference

Radunovic et al.[3] ~25dB

Existing Techniques

8

First Design: Antenna Cancellation

d d + λ/2

TX1 TX2RX

9

First Design: Antenna Cancellation

~30dB self-interference cancellation

Enables full-duplex when combined with digital (15dB) and hardware (25dB) cancellation.

d d + λ/2

TX1 TX2RX

10

Three techniques give ~70dB cancellation

• Antenna Cancellation (~30dB)

• Hardware Cancellation (~25dB)

• Digital Cancellation (~15dB)

11

Our Prototype

12

AntennaCancellation

HardwareCancellation

Digital InterferenceCancellation

Bringing It Together

QHX220

ADC

HardwareCancellation

TX Signal

AntennaCancellation

13

RX

DigitalCancellation

∑TX Samples

+-

Clean RX samples

RF

Baseband

Bringing It Together

QHX220

ADC

HardwareCancellation

TX Signal

AntennaCancellation

14

RX

DigitalCancellation

∑TX Samples

+-

Clean RX samples

RF

Baseband

Bringing It Together

QHX220

ADC

HardwareCancellation

TX Signal

AntennaCancellation

15

RX

DigitalCancellation

∑TX Samples

+-

Clean RX samples

RF

Baseband

Bringing It Together

QHX220

ADC

HardwareCancellation

TX Signal

AntennaCancellation

16

RX

DigitalCancellation

∑TX Samples

+-

Clean RX samples

RF

Baseband

-60

-55

-50

-45

-40

-35

-30

-25

0 5 10 15 20 25

RSS

I (dB

m)

Position of Receive Antenna (cm)

TX1 TX2

Only TX1 Active

Antenna Cancellation: Performance

17

-60

-55

-50

-45

-40

-35

-30

-25

0 5 10 15 20 25

RSS

I (dB

m)

Position of Receive Antenna (cm)

TX1 TX2

Only TX2 Active

Antenna Cancellation: Performance

18

Only TX1 Active

-60

-55

-50

-45

-40

-35

-30

-25

0 5 10 15 20 25

RSS

I (dB

m)

Position of Receive Antenna (cm)

TX1 TX2

Only TX1 ActiveOnly TX2 Active

Both TX1 & TX2 Active

Antenna Cancellation: Performance

19

NullPosition

-60

-55

-50

-45

-40

-35

-30

-25

0 5 10 15 20 25

RSS

I (dB

m)

Position of Receive Antenna (cm)

TX1 TX2

Only TX1 ActiveOnly TX2 Active

Both TX1 & TX2 Active

Antenna Cancellation: Performance

20

~25-30dBNull

Position

Sensitivity of Antenna Cancellation

21

Amplitude Mismatch between TX1 and TX2

Placement Errorfor RX

dB

Red

uctio

n Li

mit

(dB)

Red

uctio

n Li

mit

(dB)

Error (mm)

Sensitivity of Antenna Cancellation

22

dB

Red

uctio

n Li

mit

(dB)

Red

uctio

n Li

mit

(dB)

Error (mm)

30dB cancellation < 5% (~0.5dB) amplitude mismatch < 1mm distance mismatch

Amplitude Mismatch between TX1 and TX2

Placement Errorfor RX

Sensitivity of Antenna Cancellation

23

dB

Red

uctio

n Li

mit

(dB)

Red

uctio

n Li

mit

(dB)

Error (mm)

• Rough prototype good for 802.15.4

• More precision needed for higher power systems (802.11)

Amplitude Mismatch between TX1 and TX2

Placement Errorfor RX

Bandwidth Constraint

A λ/2 offset is precise for one frequency

24

fc

d d + λ/2

TX1 TX2RX

Bandwidth Constraint

A λ/2 offset is precise for one frequencynot for the whole bandwidth

25

fc fc+Bfc -B

d d + λ/2

TX1 TX2RX

Bandwidth Constraint

A λ/2 offset is precise for one frequencynot for the whole bandwidth

26

fc fc+Bfc -B

d d + λ/2

TX1 TX2RX

d2 d2 + λ+B/2

TX1 TX2RX

d1 d1 + λ-B/2

TX1 TX2RX

Bandwidth Constraint

27

fc fc+Bfc -B

d d + λ/2

TX1 TX2RX

d2 d2 + λ+B/2

TX1 TX2RX

d1 d1 + λ-B/2

TX1 TX2RX

WiFi (2.4G, 20MHz) => ~0.26mm precision error

A λ/2 offset is precise for one frequencynot for the whole bandwidth

Bandwidth Constraint

28

2.4 GHz

5.1 GHz

300 MHz

Bandwidth Constraint

29

2.4 GHz

5.1 GHz

300 MHz

• WiFi (2.4GHz, 20MHz): Max 47dB reduction

• Bandwidth⬆ => Cancellation⬇• Carrier Frequency⬆ => Cancellation⬆

• 802.15.4 based signaling on USRP nodes

• Two nodes at varying distances placed in an office building room and corridor

Experimental Setup

30

• Full-duplex should double aggregate throughput31

Half-Duplex :- Nodes interleave transmissions

Node 1 ➜ 2

Node 2 ➜ 1

Node 1 ➜ 2

Node 2 ➜ 1

Full-Duplex :- Nodes transmit concurrently

Median throughput 92% of ideal full-duplex

Throughput

0

0.2

0.4

0.6

0.8

1.0

0 50 100 150 200 250 300

CDF

Throughput (Kbps)

Half-Duplex Full-Duplex Ideal Full-Duplex

1.84x

32

0

0.2

0.4

0.6

0.8

1.0

0 50 100 150 200 250 300

CDF

Throughput (Kbps)

Throughput

Half-Duplex Full-Duplex Ideal Full-Duplex

1.84x

33

Performance loss at low SNR

34

First prototype gives 1.84x throughput gain with two radios compared to half-duplex with a single radio.

Limitation 1: Need 3 antennas

Limitation 2: Limited to 802.15.4

Limitation 3: Difficult farfield effects

Limitation 4: Doesn’t adapt to environment

Why Half Duplex?

TX RX TX RX

• Self-interference is millions to billions (60-90dB) stronger than received signal

• If only we could “subtract” our transmission...

Poor Man’s Subtraction

2.4 GHz

5.1 GHz

300 MHz

Cancellation using Phase OffsetSelf-Interference

Cancellation Signal

∑

Cancellation using Phase Offset

38

Self-Interference

Cancellation Signal

∑

Self-Interference

Cancellation Signal

∑

Frequency dependent, narrowband

Cancellation using Signal Inversion

39

Self-Interference

Cancellation Signal

∑

Cancellation using Signal Inversion

40

Self-Interference

Cancellation Signal

∑

Self-Interference

Cancellation Signal

∑

Frequency and bandwidth independent

Time

41

Xt +Xt/2

-Xt/2

BALUN

Second Design: Balanced to Unbalanced Conversion

Traditional Design

R

TXRF Frontend

RXRF Frontend

T R+aT

aT

1. Invert the Signal

aT R

TXRF Frontend

RXRF Frontend

2T

+T -T R+aT

balun

2. Subtract Signal

R

TXRF Frontend

RXRF Frontend

ΣR+aT-T

R+aT

balun

aT

2T

-T+T

3. Match Signals

R

TXRF Frontend

RXRF Frontend

Σv

-vT R+aT-vT

R+aT

balun

attenuator anddelay line

aT

2T

-T+T

Can Receive If v = a!

R

TXRF Frontend

RXRF Frontend

Σv

-vT

R+aTattenuator anddelay line

balun

aT

2T

-T+T

R+aT-aT

+Xt/2

• Measure wideband cancellation

• Wired experiments

• 240MHz chirp at 2.4GHz to measure response

Time

Signal Inversion Cancellation: Wideband Evaluation

47

TX

RX

Signal Inversion Cancellation Setup

∑ TX RX

Phase Offset Cancellation Setup

∑RF

Signal Splitter

Xt +Xt/2

-Xt/2

Xt

+Xt/2

λ/2 Delay

Time

48

Lower isbetter

Higher isbetter

Time

49

~50dB cancellation at 20MHz bandwidth with balun vs ~38dB with phase offset cancellation.

Significant improvement in wideband cancellation

Lower isbetter

Higher isbetter

Time

50

• From 3 antennas per node to 2 antennas

• Parameters adjustable with changing conditions

Attenuator andDelay Line

TX RX

TXRF Frontend

Xt

+Xt/2 -Xt/2

∑

RXRF Frontend

Other advantages

• Need to match self-interference power and delay

• Can’t use digital samples: saturated ADC

Adaptive RF Cancellation

51

TX RX

Attenuation & Delay

Wireless Receiver

Wireless Transmitter

RF Cancellation

TX Signal Path RX Signal Path

RF ReferenceΣ

Balun

• Need to match self-interference power and delay

• Can’t use digital samples: saturated ADC

Adaptive RF Cancellation

52

RSSI : Received Signal Strength Indicator

TX RX

Attenuation & Delay

Wireless Receiver

Wireless Transmitter

RF Cancellation

TX Signal Path RX Signal Path

RF ReferenceΣ

Balun

RSSI

• Need to match self-interference power and delay

• Can’t use digital samples: saturated ADC

Adaptive RF Cancellation

53

Use RSSI as an indicator of self-interference

TX RX

Attenuation & Delay

Wireless Receiver

Wireless Transmitter

RF Cancellation

TX Signal Path RX Signal Path

RF ReferenceΣ

Balun

RSSI

Control Feedback

54

Objective: Minimize received power

Control variables: Delay and Attenuation

TX RX

Attenuation & Delay

Wireless Receiver

Wireless Transmitter

RF Cancellation

TX Signal Path RX Signal Path

RF ReferenceΣ

Balun

RSSI

Control Feedback

55

Objective: Minimize received power

Control variables: Delay and Attenuation

56➔ Simple gradient descent approach to optimize

Objective: Minimize received power

Control variables: Delay and Attenuation

57

Typical convergence within 8-15 iterations (~1ms total)

58

Digital Interference Cancellation

TX RX

Attenuation & Delay

RF ➔ Baseband

ADC

Baseband ➔ RF

DAC

Encoder Decoder

Digital Interference Reference

RF Cancellation

TX Signal Path RX Signal Path

RF ReferenceΣ

FIR filter ∆

RSSI

Control Feedback

Channel Estimate

Balun

Bringing It All Together

59

Channel Coherence

~3dB reduction in cancellation in 1-2 seconds

~6dB reduction in <10 seconds

60

Performance

• WiFi full-duplex: with reasonable antenna separation

• Not enough for cellular full-duplex: need 20dB more

Implications

• WARPv2 boards with 2 radios

• OFDM reference code from Rice University

• 10MHz bandwidth OFDM signaling

• CSMA MAC on embedded processor

• Modified for full-duplex

Implementation

62

• CSMA/CA can’t solve this

• Schemes like RTS/CTS introduce significant overhead

APN1 N2

Current networks have hidden terminals

Mitigating Hidden Terminals

63

• CSMA/CA can’t solve this

• Schemes like RTS/CTS introduce significant overhead

APN1 N2

Since both sides transmit at the same time, no hidden terminals exist

Current networks have hidden terminals

Full Duplex solves hidden terminals APN1 N2

Mitigating Hidden Terminals

64

• CSMA/CA can’t solve this

• Schemes like RTS/CTS introduce significant overhead

APN1 N2

Since both sides transmit at the same time, no hidden terminals exist

Current networks have hidden terminals

Full Duplex solves hidden terminals APN1 N2

Mitigating Hidden Terminals

65Reduces hidden terminal losses by up to 88%

Network Congestion and WLAN Fairness

Without full-duplex:

• 1/n bandwidth for each node in network, including APDownlink Throughput = 1/n Uplink Throughput = (n-1)/n

66

Network Congestion and WLAN Fairness

Without full-duplex:

• 1/n bandwidth for each node in network, including APDownlink Throughput = 1/n Uplink Throughput = (n-1)/n

67

With full-duplex:

• AP sends and receives at the same timeDownlink Throughput = 1 Uplink Throughput = 1

Network Congestion and WLAN Fairness

68

1 AP with 4 stations without any hidden terminals

Throughput (Mbps)Throughput (Mbps)Fairness (JFI)

Upstream DownstreamFairness (JFI)

Half-Duplex 5.18 2.36 0.845

Full-Duplex 5.97 4.99 0.977

Full-duplex distributes its performance gain to improve fairness

Wide Area

(LTE)_

Apps Processor

BT

Media Processor

GPS

WLAN

UWB

DVB-H

FM/XM

All of these different radios interfere with one another

Today: limit which radios can be active at the same time

Coexistence Problem

Tomorrow: Full Duplex

Wide Area

(LTE)_

Apps Processor

BT

Media Processor

GPS

WLAN

UWB

DVB-H

FM/XM WLAN

LTE

Full duplexsignal cancellation

71

Access Point networks

Full-duplex Networking

72

Access Point networksCellular networks

Cell Basestation

Relay

Full-duplex Networking

73

Access Point networks

Multi-hop Networks

Cellular networks

Cell Basestation

Relay

Full-duplex Networking

74

Access Point networks

Multi-hop NetworksSecure Networks[1,2]

Cellular networks

Cell Basestation

Relay

Full-duplex Networking

[1] Gollakota et al. “They Can Hear Your Heartbeats: Non-Invasive Security for Implantable Medical Devices.”, in Sigcomm 2011.[2] Lee et al. “Secured Bilateral Rendezvous using Self-interference Cancellation in Wireless Networks”, in IFIP 2011.

75

Access Point networks

Multi-hop NetworksSecure Networks[1,2]

Cellular networks

Cell Basestation

Relay

Full-duplex Networking

[1] Gollakota et al. “They Can Hear Your Heartbeats: Non-Invasive Security for Implantable Medical Devices.”, in Sigcomm 2011.[2] Lee et al. “Secured Bilateral Rendezvous using Self-interference Cancellation in Wireless Networks”, in IFIP 2011.

?

Possible Dealbreakers

v != a

R

TXRF Frontend

RXRF Frontend

Σv

-vT

R+aTattenuator anddelay line

balun

aT

2T

-T+T

R+aT-aT

Frequency Selectivity

Lower isbetter

Multipath

TXRF Frontend

RXRF Frontend

Σv

attenuator anddelay line

balun

-T+T

Second path

Greatest Potential?

• Where capacity is ultra-scarce (MIMO?)?

• Fixed channels/bands?

• Drop-in improvements (incremental)?

• Channel coherence < packet time?

• We can only guess.

• We’ve built working full duplex radios‣ Breaks a fundamental assumption in wireless‣ Current limitations: transmit power: ≤ WiFi

(200mW), bandwidth (≤ 20MHz)

• How full duplex could transform wireless

• Open technical barriers and challenges

Talk In a Nutshell

Backup

83

0 10 20 30-30 -20 -10

0

10

20

30

-30

-20

-10

x axis (meters)

y ax

is (

met

ers)

84

Equal Transmit Power

Deep Nulls at 20-30m

Unequal Transmit Power

What about attenuation at intended receivers?Destructive interference can affect this signal too!

• Different transmit powers for two TX helps

0 10 20 30-30 -20 -10

0

10

20

30

-30

-20

-10

x axis (meters)

y ax

is (

met

ers)

0 10 20 30-30 -20 -10

0

10

20

30

-30

-20

-10

x axis (meters)

y ax

is (

met

ers)

85

Equal Transmit Power Unequal Transmit Power

0 10 20 30-30 -20 -10

0

10

20

30

-30

-20

-10

x axis (meters)

y ax

is (

met

ers)

-52 dBm

-58 dBm-52 dBm -100 dBm

What about attenuation at intended receivers?Destructive interference can affect this signal too!

• Different transmit powers for two TX helps