End-to-End Performance and Fairness in Multihop Wireless

Backhaul Networks

V. Gambiroza, B. Sadeghi, and E. Knightly

Rice University

Violeta Gambiroza

Backhaul Networks

InternetBackhaul network

Residential useror small business

• Backhaul networks technologies – Wireline: coax-, copper-based, fiber– Wireless

Violeta Gambiroza

Wireless Backhaul Networks TAP Networks

• Multihop wireless infrastructure– High bandwidth, good economics, deployability

• Transit Access Point (TAP)

Residential user

or small business

Ethernet

Ethernet

Ethernet

Ethernet

Wireless Backhaul

Network

Internet

Violeta Gambiroza

Fundamental Scenario

Ethernet

Ethernet

Ethernet

Ethernet

One branch of the access tree Ethernet

Ethernet

Ethernet

Ethernet

Internet

TAP1 TAP2 TAP3 TAP4

• Traffic matrix– Traffic to and from

Internet

Violeta Gambiroza

Parking Lot Scenario

• Similar to parking lot with one exit

Ethernet

Ethernet

Ethernet

Ethernet

Internet

TAP1 TAP2 TAP3 TAP4

Violeta Gambiroza

Fairness Problem

Violeta Gambiroza

Fairness Problem

Violeta Gambiroza

Fairness Problem

Goal Ensure equal shares independent of spatial location

We need multihop fairness

Violeta Gambiroza

Contributions

• Fairness reference model

• Performance study– TCP – Inter-TAP fairness algorithm

• Capacity and fairness

Ethernet

Ethernet

Ethernet

Ethernet

Wireless Backhaul

Network

Violeta Gambiroza

Outline

• Fairness reference model– Limitations of existing models– Fairness objectives– Algorithm solution space

• Performance study

• Capacity and fairness

Ethernet

Ethernet

Ethernet

Ethernet

Wireless Backhaul

Network

Violeta Gambiroza

Limitations of Existing Fairness Models: Ingress-Egress Flow Granularity

• Fairness with Ingress-Egress (IE) flow granularity– Provide fair share to each ingress-egress pair

• Ingress Aggregate (IA) flow granularity–Provide fairness on both IA and IE flow granularities -Fundamentally differentFundamentally different

• Node corresponds to TAP– TAP is small business/residence

Provide fair shares to TAPs independent of number of flows

Treat TAP’s traffic as a single aggregate

Ingress-Egress flow granularity Ingress Aggregate

flow granularity

Ingress-Egress flow granularity

Violeta Gambiroza

Our Objectives (Our Objectives vs. Classical Objectives)

Flow granularity– Ingress aggregate (IA) and Ingress-

Egress

Our Objectives Classical Objectives

– Ingress-Egress (IE)

– Bandwidth

– Wired link

Depends on fairness model

Spatial properties– Provide fair shares independent of spatial location

– Maximize spatial reuse – flows sufficiently spatially separated can transmit simultaneously

Resource– Channel access time

Medium – Multirate shared wireless

channel

Formal definition in paper

Violeta Gambiroza

Problem Statement

• Fairness reference model defined

• Distributed algorithm – Targeted at achieving shares defined by reference model

• Solution space

– Local solution – insufficient

Example: Parking lot

– Multihop solution

Flow e2e – TCP

Multihop wireless network e2e – Inter-TAP Fairness Algorithm (IFA)

Violeta Gambiroza

Outline

• Fairness reference model

• Performance study– Performance factors– TCP fairness– Inter-TAP Fairness Algorithm (IFA)

• Capacity and fairness

Ethernet

Ethernet

Ethernet

Ethernet

Wireless Backhaul

Network

Violeta Gambiroza

Performance Factors (1/2)

Factors investigated• Fairness algorithms

– Uncontrolled UDP, TCP, IFA • Media access control

– 802.11 with two-way and four-way handshake • Antenna technologies

– Omni directional, sector• Carrier sense range, multiple topologies and flow scenarios…

Other simulation specs• Channel rate constant 2 Mb/sec• 1000 byte packets

Goal

• Study end-to-end performance and fairness

Violeta Gambiroza

Performance Factors (2/2)

Well understood topologies

Increased no. of hops from destination

Reduced throughput

Increased no. of source-dest. pairs

Reduced throughput

Topology

Violeta Gambiroza

Performance Factors (2/2)

Parking lot

MU-TAP and TAP-TAP transmissions on orthogonal channels

Ethernet

Ethernet

Ethernet

Ethernet

Internet

TAP1 TAP2 TAP3 TAP4

TA(1)TA(2)

TA(3)

Topology

Violeta Gambiroza

Fairness with TCP MAC, Hidden Terminals and Information Asymmetry

320.5 320.5 320.5

1000

38.5

0

400

800

1200

1600

TA(1) TA(2) TA(3) TA(4) Total

Goo

dp

ut

[kb

/sec

]

Obj. Basic RTS/ CTS

• Idealized objective– Assumes perfect collision-

free MAC

ACK Traffic

MUs generate long lived TCP-Sack flows

Carrier sense range = transmission range Ethernet

Ethernet

Ethernet

Ethernet

TAP1 TAP2 TAP3 TAP4

Violeta Gambiroza

Fairness with TCP MAC, and Hidden Terminals and Information Asymmetry

320.5 320.5 320.5

2 20

1247

1000

38.5 48

1177

0

400

800

1200

1600

TA(1) TA(2) TA(3) TA(4) Total

Goo

dp

ut

[kb

/sec

]

Obj. Basic RTS/ CTS

ACK Traffic

MUs generate long lived TCP-Sack flows

Carrier sense range = transmission range Ethernet

Ethernet

Ethernet

Ethernet

TAP1 TAP2 TAP3 TAP4

• TAP1 and TAP2 traffic starved– Both are hidden terminals– Timeouts – significant

throughput penalty TCP generates bursts

of packets

Violeta Gambiroza

Fairness with TCP MAC, and Hidden Terminals and Information Asymmetry

320.5 320.5 320.5

2 20

1247

3 27 40.7

1000

38.5 48

1177

1058.7988

0

400

800

1200

1600

TA(1) TA(2) TA(3) TA(4) Total

Goo

dp

ut

[kb

/sec

]

Obj. Basic RTS/ CTS

• RTS/CTS exchange introduces information asymmetry [KSSK02]– TAP1 has no information

of TAP3-TAP4 trans.ACK Traffic

MUs generate long lived TCP-Sack flows

Carrier sense range = transmission range Ethernet

Ethernet

Ethernet

Ethernet

TAP1 TAP2 TAP3 TAP4

• Capacity and fairness need to be considered jointly– Total is up to 125% of

objective while two flows are starved

Violeta Gambiroza

TCP and Sector Antennas

MUs generate long lived TCP-Sack flows

TAPs use sector antennas Ethernet

Ethernet

Ethernet

Ethernet

TAP1 TAP2 TAP3 TAP4

641 641 641

2000

247313

4777

356

1386.3

730

53.3

1219

692

167

0

400

800

1200

1600

2000

2400

TA(1) TA(2) TA(3) TA(4) Total

Goo

dp

ut

[kb

/sec

]

Obj. Basic RTS/ CTS

ACK Traffic

• Impact of hidden terminals and information asymmetry mitigated

• Severe spatial bias – TAP1 traffic obtains 26% of

objective • Total goodput increased • Total goodput is 67% of the

objective

Violeta Gambiroza

Inter-TAP Fairness Algorithm (IFA)

• Idealized version of algorithm– Omniscient calculation of fair rates

Practical algorithm needs messaging

• Limit traffic rate at ingress

Violeta Gambiroza

TCP and IFA

MUs generate long lived TCP-Sack flows

Carrier sense range = transmission range

Ethernet

Ethernet

Ethernet

Ethernet

TAP1 TAP2 TAP3 TAP4

• End-to-end performance considerably improved– TAP-aggregated throughput is 59%

to 75% of the objective

• Hidden terminal problem mitigated– Contention considerably decreased– TCP cannot inject bursts of packets

320.5 320.5 320.5

1000

190 223

679

38.5 26

240

0

400

800

1200

TA(1) TA(2) TA(3) TA(4) Total

Goo

dput

[kb

/sec

]

Obj. Basic 802.11/ I FA

ACK Traffic

• Spatial bias– IFA alone cannot eliminate it

• Rates lower than the objective

Violeta Gambiroza

Inter-TAP Performance Isolation

Provide inter-TAP performance isolation independent of traffic types

ACK Traffic

326.8 326.8 326.8

993.4

209246

709.4

13 8.4

246

0

200

400

600

800

1000

TA(1) TA(2) TA(3) TA(4) Total

Goo

dput

[kb

/sec

]

Obj. Basic 802.11• TCP achieves 64% of idealized

objective, while UDP obtains 75%

• Even with balanced contention TCP reduces its rate– Having more MUs per TAP TCP

performance degraded

Each TAP has one MUTAP1: MU transmits TCP trafficTAP2 and TAP3: MU transmits UDP traffic

Ethernet

Ethernet

Ethernet

Ethernet

TAP1 TAP2 TAP3 TAP4

Violeta Gambiroza

Summary of Findings (1/2)

• Starvation of upstream flows (UDP, TCP, with or w/o RTS/CTS)– “Parking Lot” scenario results in hidden terminals and

information asymmetry

• Sector antennas and carrier sense range mitigate the hidden terminal problem– Severe spatial bias

SA: Throughput as low as 26% of targeted values CSR: Throughput as low as 34% of targeted values

• TCP able to exploit spatial reuse

Violeta Gambiroza

Summary of Findings (2/2)

• IFA approximates reference model performance

• The impact of hidden terminal problem and information asymmetry mitigated– Without any modifications to CSMA/CA

• TCP over IFA achieves 59% to 75% of idealized objective– Without any modifications to TCP

• Inter-TAP performance isolation

Violeta Gambiroza

Outline

• Fairness reference model

• Performance study

• Capacity and fairness– Maximum throughput without fairness– Fairness objectives and throughput

Ethernet

Ethernet

Ethernet

Ethernet

Wireless Backhaul

Network

Violeta Gambiroza

Problem Statement

• Compute maximum aggregate throughput– No fairness constraint

• System model– One transmission possible at time– Perfect collision-free MAC

Single contention neighborhood

Violeta Gambiroza

Aggregate Throughputwith and without Fairness Constraints

1..max11}{

F

f rl

fl

F

f

f

tf

fl

tts

Assign time-shares to maximize network throughput

Fairness constraintsFairness constraintsTemporal fairness constraint

Spatial bias removal constraint

Ingress aggregate constraint

F

f rl

fl

f

t1

1

jfji

fi CtCt frji ,

No spare time-capacitySolution:

Violeta Gambiroza

Conclusions

Fairness• Fairness reference model formally defined• Designed for multihop wireless networks

Performance study• Starvation of upstream flows• Sector antennas, larger carrier sense range, IFA

mitigate the problem• IFA approximates performance of reference model

Capacity and fairness• Need to be considered jointly

End-to-End Performance and Fairness in Multihop Wireless

Backhaul Networks

V. Gambiroza, B. Sadeghi, and E. Knightly

Rice University