Apurv Bhartia IEEE ICNP November 2015 1 Improving Spectrum Efficiency In Wireless Networks

1

Apurv BhartiaIEEE ICNP

November 2015

Improving Spectrum Efficiency

In Wireless Networks

2

• Exponential growth of data traffic

• High-end devices significantly multiply traffic• New standards (802.11n/ac, LTE) allow high data rates

Motivation: Data Traffic Surge

Can data traffic grow unchecked?

Introduction

3

• What is spectrum? Range of EM radio frequencies used to transmit data

• Usable spectrum for data transmission is limited!

Challenge: Limited Spectrum

Introduction

4

• Usable spectrum is limited• Spectrum quality is diverse Operating frequency, interference, fading, noise, etc.• Wireless medium is loss-prone Signal attenuation and multi-path propagation• Nature of wireless deployments – Uncoordinated– WLANs vs. Wireless Mesh Networks

Summary of Challenges

Introduction

5

Quantifying the Challenge

More Spectrum

60x*

Data Traffic*Cisco estimates

Imperative to focus on better utilization

of the existing spectrum

Introduction

6

Approach

Packets/Transmissions

Data/Information

Medium/Spectrum

packed in

sent on

Introduction

7

Approach

More Transmissions per unit time

More Information

per transmission

Right Spectrum

packed in

sent on

Introduction

8

Research Roadmap

O3: Sending more information per transmission

DM+: Sending more transmissions per spectrum

Selecting the right spectrum for transmission

LBRH: Coarse-grained

Smart-Fi: Fine-grained

Mobihoc ’11

ICNP ‘15

Mobicom ‘11

Under submission

Introduction

9

Talk OutlineO3: Sending more information per transmission

DM+: Sending more transmissions per spectrum




Mobihoc ‘11

ICNP ‘15

Mobicom ‘11

Under submissionPost PhD Notes

Introduction

10

Wireless Mesh Networks

More Information per Transmission

Conclusion

11

Routing in WMNs• Traditional Routing routes packets along the best path several metrics proposed (hop count, ETX, etc.)• Opportunistic routing

leverages multiple forwarders to combat loss • Routing with inter-flow coding transmit information using fewer transmissions intelligently combine (code) packets together


A

B

C

D

50%

50%

50%

50%

12

Flow 1

Flow 2

Routing Scheme No. of transmissions

Traditional Routing

Opportunistic Routing

Inter-flow Coding

Opportunistic Routing +

Inter-flow Coding

ExampleMore Information per Transmission

A


Traditional Routing


Inter-flow Coding


Inter-flow Coding

B

C

D

50%

50%

50%

50%

13

Flow 1

Flow 2

2 xmits

2 xmits

2 xmits

2 xmits

8


A


Traditional Routing


Inter-flow Coding


Inter-flow Coding

B

C

D

50%

50%

50%

50%

14

Flow 1

Flow 2

1.33 xmits

2 xmits

86.66


A


Traditional Routing


Inter-flow Coding


Inter-flow Coding

B

C

D

50%

50%

50%

50%

15

Flow 1

Flow 2

86.66

2 xmits

2 xmits

P Q

P Q

R

R=P.xor.Q


A


Traditional Routing


Inter-flow Coding


Inter-flow Coding

B

C

D

50%

50%

50%

50%

16

Flow 1

Flow 2

86.66

2 xmits

2 xmits

P Q

R

R=P.xor.Q

RPQ

R

6

2 xmits


Motivating Example (5/7)


Traditional Routing


Inter-flow Coding


Inter-flow Coding

B

C

50%

50%

50%

50%

17

Flow 1

Flow 2

A

1.33 xmits

86.66

D

1.33 xmits

6



Traditional Routing


Inter-flow Coding


Inter-flow Coding

B

C

50%

50%

50%

50%

18

Flow 1

Flow 2

A

1.33 xmits

86.66

D

1.33 xmits2

xmits

4.66

6



Traditional Routing


Inter-flow Coding


Inter-flow Coding

B

C

50%

50%

50%

50%

19

Flow 1

Flow 2

A

1.33 xmits

86.66

D

1.33 xmits2

xmits

4.66

6

Significant Performance Gains Possible!


20

OpportunisticRouting

More information, better encoding and decoding choices

Spreads information across multiple nodes

Node receives less traffic

Limited coding Hard to decode

OR + Network Coding

InterflowCoding


21

Abstraction

OpportunisticRouting

More information, better encoding and decoding choices

Spreads information across multiple nodes

UnderlayNodes

Overlay Nodes

OR + Network Coding

InterflowCoding


AB

CD

22

BD

C

Overlay Plane

Underlay Plane

Flow 1

Flow 2

O3: Key Idea

Opportunistic Routing combats loss on the underlay

A


AB

CD

23

BD

C

Overlay Plane

Underlay Plane

A

Flow 1

Flow 2

Inter-flow Coding sends more information on the overlay

O3: Key Idea



AB

CD

24

BD

C

Overlay Plane

Underlay Plane

A

Flow 1

Flow 2

Inter-flow Coding sends more information on the overlay

O3: Challenges

How to jointly optimize OR and N/W Coding?How to select overlay

and under nodes?How to build a practical protocol?



O3 Optimizatio

n Framework

Topology, demands

Input

Output

Protocol Implementation

Overlay-underlay mapping

Overlay nodes, Overlay paths

Source Rate

Limiting

Underlay forwarding

Overlay forwardin

g

25

O3: Framework

Theory

Practice


• divides packets into batches of size KIntra-coded packet• Random linear combination of all packets from a batch

Inter-coded packet• Random linear combination of 2 different flows

• Destination performs Gaussian elimination

A

R1

B

R2

αP1+ ßP2

γP1+ δP2

P1

P2

α‘Q1+ ß’Q2

γ‘Q1+ δ’Q2

Q1

Q2

λP1+ μP2 + νQ1 + ξQ2

κP1+ θP2 + ζQ1 + ρQ2

λ’P1+ μ’P2 + ν’Q1 + ξ’Q2

κ'P1+ θ’P2 + ζ’Q1 + ρ’Q2 26

λP1+ μP2

κP1+ θP2

νQ1 + ξQ2

ζQ1 + ρQ2

λ'P1+ μ’P2

κ'P1+ θ’P2

νQ1 + ξ’Q2

ζ’Q1 + ρ’Q2

O3: Packet En/Decoding


• Hierarchical architecture to combine network coding with opportunistic routing • First optimization framework to jointly optimize network coding, opportunistic routing and

rate selection• Design and implement O3, a network coding-aware opportunistic routing protocol• Extensive evaluation to show its benefits

27

O3: ContributionsMore Information per Transmission

2 4 6 8 100

0.51

1.52

2.53

3.5

UW

Number of flows

28

Throughput

(Mbps)

O3: Summary of Results

O3 Intra

MORE

COPE-RL

COPE

SPP-RL

SPP

O3

1.2-1.3x

1.5-10x

1.1-1.8x

1.4-61x

1.2-1.5x

1.9-326x

O3

Rate-limited

Traditional


2 4 6 8 12 160

0.1

0.2

0.3

0.4

0.5

MIT ROOFNET

Number of flows

29

Throughput

(Mbps)

O3: Summary of Results

O3 Intra

MORE

COPE-RL

COPE

SPP-RL

SPP

O3

1.2-1.3x

1.5-10x

1.1-1.8x

1.4-61x

1.2-1.5x

1.9-326x

O3Rate-limited

Traditional


30





Mobihoc 2011

Mobicom 2011

Under submission

Post PhD Notes


31





Mobihoc 2011

Mobicom 2011

Under submission

Post PhD Notes

Right Spectrum for Transmission

32

Spectrum Selection: Approach

Selecting the right spectrum

for transmission

Coarse-grained

Fine-grainedHow to use

the selected channel?

How to select the

best channel?


33

Spectrum Selection: Approach


for transmission

Coarse-grained



How to select the

best channel?Smart-Fi


34

Channel quality is uniform

All symbols are equal

Significant frequency diversity exists

Not all symbols are equalHeader vs. payload symbolsData symbols vs. FEC symbols (Systematic FEC)Subject vs. Background symbols

Smart-Fi: Motivation

Smart-Fi: Fine Grained Spectrum Selection

Existing Wi-Fi Protocols

35

1 10 20 30-5

0

5

10

15

20

25

30

Sig

nal to

Nois

e

Rati

o (d

B)

Ch

an

nel Q

uality

Subcarriers within a 20MHz channel

Frequency Diversity

Expected?


36

1 10 20 30-5

0

5

10

15

20

25

30

Subcarriers within a 20MHz channel

Frequency Diversity

Frequency selective variability, narrow-band interference

Sig

nal to

Nois

e

Rati

o (d

B)

Ch

an

nel Q

uality

Actual!


37

802.11n Up to 40 MHz

802.11ac Up to 160 MHz

Ultra Wideband 100s of MHz to GHz

Frequency diversity increases with wider channels!

Move to Wider Channels


38

Channel Quality Trace Analysis

Unified Approach

Evaluation

Smart-Fi: Approach

Fine-grained channel quality information

Differential

Protection

Smart Symbol

Mapping

Enhanced Error

Correction

Smart-Fi


39

• Avoid long runs of low reliability bits but assumes all subcarriers are equal all bits are equal

• Arranges bits in a non-contiguous wayImproves performance of FEC codesStandard 2-step permutation process

OFDM Standard Interleaving


40

• Map important symbols to reliable subcarriersMapping should maximize throughput

ProblemGiven a set of subcarriers, determine symbol-subcarrier mapping that maximizes the expected received payload

i.e.

• Non-linear utility functionOptimal solution is challengingWe develop several heuristics …

correctly received data bits in FEC group

Smart Symbol Mapping (1/2)


Smart Header/DataSubcarriers ordered by SNR

Data

FEC

Data

FEC

Smart DataFEC

Header Payload

Smart Header

Header

Payload41

Header

Payload

Data FEC

Header(Data)

Payload(Data)

Header(FEC)

Payload(FEC)

High Low SNR

Data

Smart Symbol Mapping (2/2)


42

• One PHY-layer data rate might not work for all subcarriersPer subcarrier modulation and PHY-layer FEC? [e.g. FARA]Map same FEC group symbols to nearby subcarriers bursty lossesSignificant signaling and processing overheadNot available in commodity hardware

• Benefits of MAC-layer FECProtection based on symbol importanceMore fine-grained than PHY-layer FECEasily deployable on commodity hardware

Differential Protection (1/2)


43

• Maximize throughput by selectively adding MAC FEC

• Challenge: Search space becomes larger!How much MAC FEC to add?How to split MAC FEC to differentially protect PHY-layer symbols?What FEC group size to use at the MAC layer?

MAC-layer FEC

FEC Group

Redundancy Symbols

Data Symbols

PHY-layer Frame

Differential Protection (2/2)


44

Smart-Fi

Differential

Protection

Smart Symbol

Mapping

Enhanced Error

Correction

Unified Approach

1.6-5x

1.7-1.8x

1.1-3x

2.6-7.6x

Smart-Fi: Summary of Results


45

Selecting the Right Spectrum


for transmission

Coarse-grained



How to select the

best channel?Smart-Fi


46

Selecting the Right Spectrum


for transmission

Coarse-grained



How to select the

best channel?Smart-Fi LBRH

LBRH: Coarse Grained Spectrum Selection

47

• FCC released spectrum in 50-700MHz spectrum Spectrum released after analog-to-digital TV transitionFree for use by unlicensed devices

• Potential for increased wireless coverage

Whitespaces: An Introduction


48

Seattle, WA

• Allow wide-area coverage Uncoordinated deployment of APs over wide area• Static spectrum sharing is difficult• Fragmented spectrum with variable sized blocks

Austin, TX

Whitespaces: Challenges


49

Desired PropertiesStati

cLCCS, MinM

ax

SSCH,

MaxChop

mCham

[Sigcomm’09]

Low Low Medi

umMedium

High

Low Low Medium

Medium

High

DistributedFairn

essUtilizationMobil

ityWide AreaFrequency

Diversity

Channel Bonding

Fragmented Spectrum

2 radios

1 radi

o


50

1 3 7-10

5 MHz 20 MHz5 MHz

A

B

C

A

CB

A

B

C

LBRH: Motivating Example-1 Spect

rumTopology


51

A

CB

20MHz

5MHz

5MHz

5MHz

20MHz

5MHz

20MHz

5MHz

5MHzA

B

C

Time

A

B

C

Time

6.7MHz

mCham (Optimal) Hopping

Utilization

LBRH: Motivating Example-1

1 3 7-10

5 MHz 20 MHz5 MHzSpectrum

Topology

t1

t2

t1

t2

20MHz

10MHz

6.7MHz10M

Hz6.7MHz


52

1 3-5 7-10

15 MHz 20 MHz5 MHz

A

B

C

A

CB

A

B

C

LBRH: Motivating Example-2 Spect

rumTopology


53

A

CB

20MHz

5MHz

5MHz

15MHz

20MHz

15MHz

20MHz

5MHz

15MHzA

B

C

Time

A

B

C

Time

15MHz

20/2=10MHz

mCham (Optimal) Hopping

Utilization

Fairness

LBRH: Motivating Example-2

1 3-5 7-10


Topology

t1

t2

t1

t2

20MHz

20/2=10MHz


54

LBRH: Key Idea-IA

CB 1 3 7-10


Topology

mCham

Random Hopping

Optimal

Utilization (MHz)

Channel Hopping is essential

20

30

8.3


55

• Accurate prediction of channel is important• Overhearing does not always work

Interference range is greater (2X) than transmission range

• Predicting throughput analytically is a hard problem!

A

CB1 3-5 7-10

15 MHz 20 MHz5 MHz

LBRH: Key Idea-II

? ?

Unused Spectrum

A

CB

Participatory Measurement is critical


56

Good channels – longer durationBad channels –

shorter duration

Channel Hopping with a ‘bias’

Participatory Measurement using a ‘leaky

bucket’

Leaky Bucket Random Hopping


57

Random Hopping

set of Maximally Bonded Channels MHz {1, 3-5, 7-10}

LBRH: Working

1 3 4 5 7

15M 20M5M

Every

On each successful transmission, Leaky

Bucket

8910


58

Random Hopping


LBRH: Working

1 3 4 5 7

15M 20M5M

Every


Bucket

8910

Fill the bucket with a random number of tokens


59

Random Hopping


LBRH: Working

1 3 4 5 7

15M 20M5M

Every


Bucket

8910

Randomly hop to a channel from


60

Random Hopping


LBRH: Working

1 3 4 5 7

15M 20M5M

Every


Bucket

8910

Record the current bandwidth of the channel


61

Random Hopping


LBRH: Working

1 3 4 5 7

15M 20M5M

Every


Bucket

8910

Every time slot, debit tokens from the bucket

Debit Tokens


62

Random Hopping


LBRH: Working

1 3 4 5 7

15M 20M5M

Every


Bucket

8910

On every successful transmission, credit bucket

Debit TokensCredit Tokens


63

Random Hopping


LBRH: Working

1 3 4 5 7

15M 20M5M

Every


Bucket

8910

Whenever bucket becomes empty, re-iterate


64

LBRH: Intuition

1 3-5 7-10

15 MHz 20 MHz5 MHzA

B C

20MHz

15MHz

5MHz 20MHz

5MHz

20MHz

15MHz

20MHz

A

B

CLBRH


65

LBRH: Intuition

1 3-5 7-10

15 MHz 20 MHz5 MHzA

B C

20MHz

15MHz

5MHz

5MHz

20MHz15MHz

20MHz

A

B

C

5MHz

15MHz

Sub-optimal

Random Hopping


66

LBRH: Working

1 3-5 7-10

15 MHz 20 MHz5 MHzA

B C

20/2=10MHz

15MHz

20/2=10MHzA

B

CmCha

m


67

LBRH: Intuition

Configurations

Utilization (MHz)

{5,15,20}40

{15,20,20}35

{5,20,20}35

{15,15,20}

35

LBRH 98.8 0.5 0.4 0.2mCham 0.0 100.0 0.0 0.0Random Hopping

25.1 24.9 24.8 25.2LBRH biases towards high utilization configurations

Sub-optimalOptimal


68

LBRH: Revisiting Example

A

CB 1 3 7-10


Topology

mCham

RH Optimal

Utilization (MHz)

20

30

8.3


69

LBRH: Revisiting Example

A

CB 1 3 7-10


Topology

mCham

LBRH Optimal

Utilization (MHz)

20

30

29.5

LBRH is close to optimal in utilization and fairness


70

• Simulated on QualNet for large-scale validationPacket-level network simulator (universally accepted)Actual channel availability using Google Spectrum DatabaseVaried topology extensively - 20 APs and 20 clients with 10 random topologies

- Vary the interference degree of each AP Varied traffic demands and patterns (CBR, HTTP, etc.)• Implemented on SORA testbed Software radio platform by Microsoft Research Asia

LBRH: Evaluation Methodology


71

LBRH: Summary of Results

LBRH outperforms mCham (1.4-1.6x), RH (1.6-1.9x)

Run 1 Run 2 Run 3 Run 4 Run 50

10

20

30

40

50

60

Tota

l Thr

ough

put (

Mbp

s)


72

0 1 2 30.1

0.4

0.7

1

0 1 2 30.1

0.4

0.7

1

Data Rate (Mbps) Data Rate (Mbps)

CDF

CDF


RH

LBRH

mCham

RH

LBRH

mCham

AUSTIN, TX ITHACA, NY

Improves lowest 10% of the flows by upto 4X


73

0 1 2 30.1

0.4

0.7

1

0 1 2 30.1

0.4

0.7

1


CDF

CDF


RH

LBRH

mCham

RH

LBRH

mCham


Fairness is very close to optimal

mCham

RH LBRH

Optimal

0.83 0.99 0.99 1.00


74

0 1 2 30.1

0.4

0.7

1

0 1 2 30.1

0.4

0.7

1


CDF

CDF


RH

LBRH

mCham

RH

LBRH

mCham


Benefits improve as channels increase (e.g. Ithaca)


75


LBRH

LBRH converges to an equilibrium quickly

0 9 18 27 36 45 54 63 72 81 90 990

10

20

30

40

50

60

Time (seconds)

Thro

ughp

ut (M

bps)

All the nodes are active


76


LBRH adapts well to topology changes

3 nodes

4 nodes

5 nodes

Number of channels

Flow

rate

s [M

bps]

10

20

30

40

1 2 3 1 2 3 4 1 2 3 4


77

• Selecting the right spectrum is criticalUse the selected channel efficientlySelect the appropriate channel for transmission

• Smart-Fi: harnesses frequency diversity of the channelComplementary techniques to improve performanceBenefits: upto 6.6x in fixed rate, 134% in auto-rate

• LBRH: distributed spectrum sharing mechanism Enables high-throughput, fair-sharing in WWANs Benefits: upto 55% over state-of-art, almost optimally fair

Spectrum Selection: Summary


78





Mobihoc 2011

Mobicom 2011

Under submission

Post PhD Notes

Conclusion

79

More information per packet

More transmissions per unit time

Right spectrum selection

Network Coding with Opportunistic Routing

Distributed MIMO

Smart-Fi (fine-grained), LBRH (coarse-grained)

(4.4x * 1.5x)

1.4x2.2x

Discussion60x

Data

Building block-1

Building block-2

Building block-3

Building block-4

High Throughput Wireless Systems

Spectrum Utilization

20x

Conclusion

80





Mobihoc 2011

Mobicom 2011

Under submission

Post PhD Notes

Post PhD Notes

81





Mobihoc 2011

Mobicom 2011

Under submission

Post PhD Notes

Post PhD Notes

82

• Wanted to work on something for immediate impactsolve real issues at hand (and not 10 yrs. from now)work on scale

• Build systems which (hopefully) millions would use• Move to the SF Bay area (where action happens!)• More stable lifestyle• Not worry about funding, grants, etc.

Why I chose Industry?

Post PhD Notes

83

• Founded by MIT PhD students• Enterprise cloud-managed wireless networks• Multiple product lines

Access points , Switches and Security Appliances

• More than 100k networks deployed worldwide• Full stack development• Around 100 engineers, often collaborate in small teams• Acquired by Cisco in Nov. 2012

What Meraki doesPost PhD Notes

84

• Optimizing performance in real wireless networksChannel assignment, scheduling, aggregation, etc.Lots of real data (helps in papers, analysis)Often involves lot of driver-level work

• Accepted paper at Sigcomm’15Large Scale Measurements of Wireless Network Behavior

• Quality of Service (QoS) improvementsScheduling for fairness, and scalabilityMassive drive towards newer standards (e.g. 802.11ac)

So far … Post PhD Notes

85

[email protected]

FIN

86

[email protected]

FIN

87

• Location flexibility?• Finanical factors? (grants, funding)• Work impact? (papers vs. patents)• Ownership? (self vs. company)• Recognition• Time flexibility?

Academia vs. Industry

Post PhD Notes

88

89

Dissertation My Work

LBRH*Frequency Diversity in Wi-Fi (Mobicom’11)CRMA (Mobicom’11)

Right Spectrum for TransmissionDM+*

Multi-point MIMO (Infocom’13)

More Transmission per Spectrum

FRJ (IWQoS’09)Energy-aware Rate Adaptation (Mobihoc’13)Smart Retransmissions*

Minimizing Transmission Time, Rate Adaptation

Wireless Display (on Windows Phone)+

Clean Slate Architecture for Internet (SDN)+

Other Works

O3 (Mobihoc’11) More Information per Transmission

*Under Submission +Patents filed/issued

WiFi-XL: A license-Free Wireless LAN (In submission)Apurv Bhartia, Mahanth Gowda, Krishna Chintalapudi, Bozidar Radunovic, Ramachandran Ramjee, Lili Qiu and Romit Roy Choudhury.

DM+: Embracing Distributed MIMO in Wireless Mesh Networks (In submission)Apurv Bhartia, Yi-Chao Chen, George Nychis and Lili Qiu.

Smart Retransmissions (In submission)M. Owais Khan, Apurv Bhartia and Lili Qiu.

Model-Driven Energy-Aware Rate AdaptationM. Owais Khan, Vacha Dave, Yi-Chao Chen, Oliver Jensen, Lili Qiu, Apurv Bhartia and Swati Rallapalli. ACM MobiHoc, Bangalore, India, July 2013.

Multi-point to Multi-point MIMO in Wireless LANsSangki Yun, Lili Qiu and Apurv Bhartia. IEEE Infocom [Mini Conference], Turin, Italy, April 2013.

Harnessing Frequency Diversity in Wi-Fi NetworksApurv Bhartia, Yi-Chao Chen, Swati Rallapalli and Lili Qiu. ACM MobiCom, Las Vegas, NV, USA, Sept 2011

CRMA: Collision-Resistant Multiple Access (Best Paper Nominee)Tianji Li, Mi Kyung Han, Apurv Bhartia, Lili Qiu, Eric Rozner, Yin Zhang and Brad Zarikoff. ACM MobiCom, Las Vegas, NV, USA, Sept 2011

O3: Optimized Overlay-Based Opportunistic RoutingMi Kyung Han, Apurv Bhartia, Lili Qiu and Eric Rozner.ACM MobiHoc, Paris, France, May 2011. Fast Resilient Jumbo Frames in Wireless LANsAnand Padmanabha Iyer, Gaurav Deshpande, Eric Rozner, Apurv Bhartia and Lili Qiu.IEEE IWQoS, Charleston, SC, June 2009.

90

Publications

91

More information per packet

More transmissions per unit time

Right spectrum selection

Network Coding with Opportunistic Routing

Distributed MIMO

Smart-Fi (fine-grained), LBRH (coarse-grained)

(3.4x * 1.5x)

1.4x2.1x

“The Vision”60x

Data

Building block-1

Building block-2

Building block-3

Building block-4

High Throughput Wireless Systems

Spectrum Utilization

15x

Documents

Apurv Bhartia IEEE ICNP November 2015 1 Improving Spectrum Efficiency In Wireless Networks