BBN: Throughput Scaling in Dense Enterprise

WLANs with Blind Beamforming and Nulling

Wenjie Zhou (Co-Primary Author), Tarun Bansal (Co-Primary Author), Prasun Sinha and Kannan Srinivasan

The Ohio State University

Changes in Uplink Traffic

2

Cloud Computing

Online Gaming

Sensor Data Upload

Code Offloading VoIP,

Video Chat

Traditionally, WLAN traffic: • downlink heavy• less attention to uplink traffic

Recently, uplink traffic increased rapidly : • mobile applications

Can we scale the uplink throughput with the number of clients?

Network MIMO

Huge bandwidth consumption

C2C1 C3

Exchange raw samplesAP1 AP2 AP3

[1] Rahul, H., Kumar, S., and Katabi, D. MegaMIMO: Scaling Wireless Capacity with User Demand. In Proc. of ACM SIGCOMM 2012.

MegaMIMO1

Does not apply to uplink :Clients do not share a backbone network

[1] Cadambe, V. R., and Jafar, S. A. Interference Alignment and the Degrees of Freedom for the K User Interference Channel. IEEE Transactions on Information Theory (2008).

Interference Alignment1

• 4 packets, 3 slots• Enough time slots, everyone gets half the cake • Exponential slots of transmissions, not suitable for mobile clients• Heavy coordination between clients

C2

C1

C3

AP1

AP2

AP3

Existing interference alignment and beamforming techniques are not suitable to mobile uplink traffic.

How can we bring the benefits of beamforming to uplink traffic?

AP Density in Enterprise WLANs

8

50 70 90 110 130 150 170 1900

0.25

0.5

0.75

1

Number of Access Points (APs)

CDF

(140,0.5)

BBN leverages the high density of access points

Single Collision Domain

C1 C2 C3

x1 x2 x3

AP1 AP2

AP3 AP4

Switch

Omniscient TDMA

Time Slot: 1Time Slot: 2Time Slot: 3

Three Packets received in Three Slots Only one AP is in use 9

10

h(1)12x1 + h(1)

22x2 + h(1)32x3h(1)

11x1 + h(1)21x2 + h(1)

31x3

Blind Beamforming and NullingSingle Collision Domain

Time Slot: 1

C1 C2 C3

x1 x2 x3

AP1 AP2

AP3 AP4

Switchh(1)

13x1 + h(1)23x2 + h(1)

33x3 h(1)14x1 + h(1)

24x2 + h(1)34x3

h(1)13 h(1)

23h(1)

33

11

Receives:a11x1 + s1h(1)

21x2 + s1h(1)31x3

Receives:a12x1 + a22x2 + a32x3

Transmits:

v4 (h(1)14x1 + h(1)

24x2 + h(1)34x3)

Transmits:

(h(1)13x1 + h(1)

23x2 + h(1)33x3)

Time Slot: 2

Blind Beamforming and NullingSingle Collision Domain

AP1 AP2

AP3 AP4

Switch

v3

12

AP1 AP2

AP3 AP4

Switch

Slot 2: a11x1 + s1h(1)21x2 + s1h(1)

31x3 Slot 2: a12x1 + a22x2 + a32x3

Slot 1: h(1)11x1 + h(1)

21x2 + h(1)31x3 Slot 1: h(1)

12x1 + h(1)22x2 + h(1)

32x3

Three Packets received in Two Slots

Blind Beamforming and NullingSingle Collision Domain

(s1h(1)11-a11)x1

Slot 2: a11x1 + s1h(1)21x2 + s1h(1)

31x3

Number of APs Required

• In a network with APs, APs in BBN can

receive N uplink packets in two slots

• 3 clients, 4 APs

• 4 clients, 7 APs

• 10 clients, 46 APs

13

2

22 NN

Throughput Improvement

• Previous Example Topology– APs in BBN receive three packets in two slots: an

improvement of 50%

• General Topology– Uplink throughput in BBN scales with the number of

clients (N/2 packets per slot). – Half of the cake as in Interference Alignment

• Always two slots• No coordination between clients

14

BBN Highlights

• Leverages the high density of access points • All computation and design complexity shifted to

APs • APs only need to exchange decoded packets over

the backbone instead of raw samples

15

Further Optimizations to Improve SNR

• Which subset of APs act as transmitters and which subset as receivers?

• Which AP decodes which packet?

C1 C2 C3

AP1AP2

AP3

AP4

Switch

16

BBN Approach: xi is decoded at the APj where it is expected to have highest SNR

Transmitters

Receiversx1 x2, x3

Challenge 1/4: Synchronization of APs

• To perform accurate beamforming, APs need to be tightly synchronized with each other

• Solution: – SourceSync (Rahul et al., SIGCOMM 2010):

synchronizes APs within a single collision domain – Vidyut (Yenamandra et al., SIGCOMM 2014):

uses power line to synchronize APs in the same building

17

Challenge 2/4 : MultiCollision Domain

• Not all APs may be able to hear each other directly

• Solution: Make smaller groups where all APs in a single group can hear each other.

18

19

Distributed System

Group Head

Group Head

• Within a group, all APs can hear each other• When one group is communicating, neighboring groups

remain silent

Challenge 3/4 : Inconsistency in the AP density

• Number of APs may be less than

• Solution: Appropriate MAC layer algorithm that restricts the number of participating clients

2

22 NN

20

21

Uplink

Poll Approve A, B and C

Keep Silent – Allow neighboring groups to transmit

Downlink Uplink

....... ....... .......

Time

Notification Period

Time Slot 1 Time Slot 2

Uplink

MAC Timeline

Compute pre-coding vectors in the background

C1 C2 C3

x1 x2 x3

AP1 AP2

AP4 AP5

Switch

Challenge 4/4 : Robustness• Nulling is not always perfect.

x1, x2 , x3x1

Decoding Error

Can’t Subtract x1

22

C1 C2 C3

x1 x2 x3

AP1 AP2

AP3 AP4

Switch

Challenge 4/4 : Robustness• What if we have extra APs

AP5

AP6AP7

x1, x2 , x3x1 x1

23

Experiments

24

C1 C2

x1

x2, x3AP1 AP2

AP3 AP4

Switch

Intended Signal = x1

Interference from x2, x3

x2

C2

x3

USRP N210

25

Throughput

BBN provides 1.48x throughput compared to TDMA

1.48X

Trace-Driven Simulation• Over multiple collision domains (divided into groups)

• Field Size: 500m X 500m

• Number of clients: 1000

• Vary the number of APs

• Residual interference distribution from experiment

• Other algorithms simulated– Omniscient TDMA– IEEE 802.11 26

27

• 2000 APs• 4.6X throughput

gain• ~76 APs near each

client

Throughput

BBN

Fairness

28

BBN achieves higher fairness• Beamforming increased SINR of clients that are far

away

BBN

29

Summary and Future Work• BBN leverages the high density of APs to scale the uplink

throughput for single antenna systems– Throughput scales linearly with the number of clients– All computational and design complexity shifted to APs

• Future Work– Coexist with legacy network

– Data rate selection

Thank you

Backup Slides

30

Long Term Results

OctoClock-G

Frequency Accuracy w/ out : 25 ppb Frequency Accuracy with GPS Lock : <1 ppb PPS Accuracy with GPS Lock : 50 ns

Vidyut : approximately 225 ns

Multiple Antenna AP

• Assume each AP has K antennas• For N clients, APs required• For M APs, clients

K

KNN

2

22

MK2

Estimate SNR of C1 at AP2

SNR of C1 at AP2 is low

C1

AP1 AP2

AP3

Switch

AP4

34

No path with high

SNR

Estimate SNR of C1 at AP1

SNR of C1 at AP1 is high

C1

AP1 AP2

AP3 AP4

Switch

35

One path with high

SNR

• C1 should be decoded by AP1

• AP1 should act as a receiver in slot 2

3

2

1

00

00

00

)( Clientsby dTransmitte Data

x

x

x

X

37363534

27262524

17161514 )( APs Receiving toClients from Channel

hhhh

hhhh

hhhh

H

Blind Nulling in BBN

36

4

3

2

1

000

000

000

000

)( Vectors Precoding

v

v

v

v

V

73'

72'

71'

63'

62'

61'

53'

52'

51'

43'

42'

41'

)( APs Among Channel

hhh

hhh

hhh

hhh

H '

AP1 :

C1 : AC1 Packet 1

C2 : AC2 Packet 2

C3 : AC3 Packet 3

IACS IACS IACS

Approve

MAC Layer: Phase 1

37

AP1 :

AP2:

AC3AP3 :

AP4 : v4* Samples4

BIFSAC1

AC2

SIFS SIFS SIFS

AP5 : v5* Samples5

AP6: v6* Samples6

AP7 : v7* Samples7

MAC Layer: Phase 2

38

Experiments Setup

• Performed using USRP N210 Radio

• Testbed of 4 APs and 3 clients

• Modulation Scheme: OFDM with BPSK

• Channel: Central Frequency 400 Mhz, Bandwith set to 500 KHz

39

Existing Schemes

• Interference Alignment– Existing IA schemes perform alignment over exponential number of time

slots [Cadambe et al., IEEE Transactions on Information Theory 2007]

• MU-MIMO (Multi User MIMO)– Requires transmitters to exchange each other’s data before transmission

• MU-MIMO (Multi User MIMO) in EWLAN– All APs together act as a single AP with multiple antennas– Requires APs to exchange samples over the backbone which is cost-

prohibitive [Gollakota et al., SIGCOMM 2009; Gowda et al., INFOCOM 2013]

40

Existing Schemes

• Interference Alignment– Existing IA schemes require each transmitter to transmit

exponential amount of data [Cadambe et al., IEEE Transactions on Information Theory 2007]

• MU-MIMO– All APs together act as a single AP with multiple antennas– Requires APs to exchange samples over the backbone

which is cost-prohibitive [Gollakota et al., SIGCOMM 2009]

41

Related Work (contd.)

• Interference Alignment– Existing IA schemes work over exponential number of time slots

[Cadambe et al., IEEE Transactions on Information Theory 2007]

– Or, work only for downlink [Suh et al., IEEE Transactions on Communications 2011]

– Or, require multiple antennas at clients [Gollakota et al., SIGCOMM 2009]

– Or, require APs to exchange samples over backbone [Annapureddy et al., IEEE Transactions on Information Theory 2012] 42

Related Work

• Backbone Usage–MegaMIMO (Rahul et al., SIGCOMM 2012):

Works only for downlink

– Symphony (Bansal et al., MobiCom 2013): Works only in multiple collision domain

43

Related Work (contd.)

• Wireless Relays– Use special relay nodes to assist high speed

communication between specific transmitters and receivers

– Existing algorithms do not make use of the backbone

– BBN leverages the backbone to improve throughput

– BBN can extend to multiple rounds to decode packets with low SNR

44

BBN Highlights

• Leverages the high density of access points • Uplink throughput scales with the number of

clients in the network• All computational and design complexity shifted

to APs• APs only need to exchange decoded packets over

the backbone

45

C1 C2 C3

x1 x2 x3

AP1 AP2 AP3

AP4 AP5

Switch

AP6 AP7

Example Topology: What we ideally want

46

x2x1 x3

Works! But can we make the requirements less strict?

Matching in BBN

47

C1

C2

C3

AP1

AP2

AP3

AP4

AP5

AP6

AP7

Edge Weight = Expected SINR of C2 at AP3

Find the Maximum Weight Matching

• Which AP decodes which packet.• Which AP transmits in the second slot.

Number of APs Required: Example Topology

• Two packets (x2 and x3) need to be nulled at AP1

• One packet (x3) needs to be nulled at AP2

• Three transmitting APs required• Guarantee non degenerate solution: Four APs required

48

x1 x2x3

AP Density in Enterprise WLANs

49

60 80 100 120 140 160 180 2000

0.25

0.5

0.75

1

CDF of Number of APs Observed

Number of Access Points (APs)

CDF

CDF of number of APs observed (Measurements conducted at Ohio State

University campus)

Can we leverage the high density of APs to scale the uplink throughput?

Enterprise Wireless LAN

50

AP AP AP

AP AP AP

Internet

BBN Overview

• Leverages the high density of access points• Uplink throughput scales with the number of

clients in the network

– Schedule length: Two Slots• First slot: Clients transmit• Second slot: APs perform blind nulling

– APs only need to exchange decoded packets over the backbone

51

52

Contents

• BBN Design• Experiments• Simulations• Challenges• Related Work• Conclusion

C1 C2 C3

x1 x2 x3

AP1 AP2 AP3

AP4 AP5

Switch

AP6 AP7

Example Topology (Single Collision Domain) with BBN

53

Time Slot: 1

C1 C2 C3

x1 x2 x3

AP1 AP2 AP3

AP4 AP5

Switch

AP6 AP7

Simultaneous Nulling Goal

54

Receive: x1, x2

Null: x3

Receive: x1

Null: x2, x3

Receive: x1,, x2, x3

C1 C2 C3

x1 x2 x3

AP1 AP2 AP3

AP4 AP5

Switch

AP6 AP7

Example Topology (Single Collision Domain) with BBN

55

Time Slot: 1

h14x1 + h24x2 + h34x3

h15x1 + h25x2 + h35x3

... ...

Time Slot: 2

v4 * (h14x1 + h24x2 + h34x3 ) v5 * (h15x1 +

h25x2 + h35x3 )

v6 * (...) v7 * (...)

a11x1 a12x1 + a22x2 a13x1 + a23x2 + a33x3

AP1 AP2 AP3Sw

itch

Example Topology (Single Collision Domain) with BBN

56

Time Slot: 2a11x1 a12x1 + a22x2 a13x1 + a23x2 + a33x3

- a12x1

= a22x2- a13x1

- a23x2

= a33x3

Three Packets received in Two Slots

Blind Nulling in BBN (Contd.)

• What we want: Blindly null x2 and x3 at AP1; and, blindly null x3 at AP2

• Assuming nulling and packet cancellation is perfect, Compute V such that

57

000

000

000'VHH

MAC Layer

58

Uplink

Poll Approve A, B and C

Keep Silent – Allow neighboring groups to transmit

Downlink Uplink

....... ....... .......

Time

Contention Period

Phase I Phase II Phase III

RSS Without Blind Beamforming and Nulling

59

-55

-45

-35

-25

-15

Signal Strength at AP1 Before Cancellation (in dB)

Interference Intended Signal

-8dB SINR

RSS After Blind Beamforming

60

-55

-45

-35

-25

-15

Signal Strength at AP1 After Cancellation (in dB)

Interference Intended Signal

13dB SINR

• Proper precoding increases SINR by 21 dB• Blind Beamforming and Nulling is practical

61

How to boost uplink throughput?

MIMO?

netwrok-MIMO?

Smartphones are small

High bandwidth consumption

Interference Alignment?

Smartphones are mobile