Predictable Performance Predictable Performance Optimization for Wireless NetworksOptimization for Wireless Networks

Lili Qiu Lili Qiu University of Texas at AustinUniversity of Texas at Austin

lili@cs.utexas.edulili@cs.utexas.edu

Joint work with Joint work with

Yi Li, Yin Zhang, Ratul Mahajan, and Eric RoznerYi Li, Yin Zhang, Ratul Mahajan, and Eric Rozner

ACM SIGCOMM 2008ACM SIGCOMM 2008August 21, 2008August 21, 2008

2

MotivationMotivation• Wireless networks are becoming ubiquitous• Managing wireless networks is hard

• Our goal: develop systematic techniques to optimize wireless performance

Predict if given sending rates are achievable

Perform what-if analysis

Optimize sending rates for different objectives

Wireline Wireless

3

0 200 400 600 800

1000 1200 1400

0 1000 2000 3000 4000 5000

Thro

ughp

ut (K

bps)

Sending rate (Kbps)

bad-goodgood-bad

Unpredictability of wireless networksUnpredictability of wireless networks

Need predictable wireless performance optimization.

S

S

R

R

D

DSS

50%

100%

100%

50%

bad-good

good-bad

4

Model-driven optimization frameworkModel-driven optimization framework

Network measureme

nt

Network model

Optimization

Traffic demands prescribed

flow rates

Performance objectives: - Maximize fairness, total throughput, …

Routing

5

Existing models are insufficientExisting models are insufficient• Asymptotic performance bounds [GP00,LB+01,GT01,GV02]

– Cannot model any specific networks

• Conflict graph based model [JPPQ03]– Assumes perfect scheduling and overestimates 802.11

performance– Requires an exponential number of constraints

• 802.11 DCF models [Bianchi00,KA+05,GLC06,GSK05 QZWH+07,KDG07]

– Not general: restricted topologies or traffic demands– Cannot be easily incorporated into optimization

procedure

Need a better 802.11 network model for optimization.

6

Our network modelOur network model• Provide a compact characterization of feasible sol

ution space to facilitate optimization• Simple: O(N) constraints for N links• Flexible and accurate

– Handle asymmetric link loss rate– Handle asymmetric interference– Handle hidden terminals– Handle heterogeneous, multihop traffic demands

Network measureme

nt

Network model

Throughput constraints

Loss rate constraints

Sending rate constraints

7

Throughput constraintsThroughput constraints• Divide time into variable-length slot

(VLS)– 3 types of slots:

idle slot, transmission slot, deferral slot

j ijjjijiislotj

iiii TDTT

pEPg

)1(

)1(

Expected payloadtransmission time

Probability of starting tx in a slot

Success probability

Expected duration of a variable-length slot

8

Loss rate constraintsLoss rate constraints • Inherent and collision loss are independent • Inherent loss

– Based on one-sender broadcast measurement

• Collision loss– Synchronous loss

• Two senders can carrier sense each other• Occur when two transmissions start at the same time

– Asynchronous loss• At least one sender cannot carrier sense the other• Occur when two transmissions overlap

9

Sending rate feasibility Sending rate feasibility constraintsconstraints

• 802.11 unicast

– Random backoff interval uniformly chosen [0,CW]

– CW doubles after a failed transmission until CWmax, and restores to CWmin after a successful transmission or when max retry count is reached

– CW(pi): the expected contention window size under packet loss rate pi [Bianchi00]

• Sending rate feasibility constraints

2/)(1

10

ipCWi

DIFS Data TransmissionRandomBackoff

ACKTransmission

SIFS

10

Extensions to the basic modelExtensions to the basic model• RTS/CTS

– Add RTS and CTS delay to VLS duration– Add RTS and CTS related loss to loss rate constraints

• Multihop traffic demands– Link load routing matrix e2e demand– Routing matrix gives the fraction of each e2e demand th

at traverses each link• TCP traffic

– Update the routing matrix:

where reflects the size & frequency of TCP ACKsackdataTCP RRR

11

Model-driven optimization frameworkModel-driven optimization framework

Network measureme

nt

Network model

Optimization

Traffic demands prescribed

flow rates

Performance objectives: - Maximize fairness, total throughput, …

Routing

12

Flow throughput feasibility testingFlow throughput feasibility testing• Test if given flow throughput are achievable• Challenge: strong interdependency• Our approach: iterative procedure

Initializeτ= 0 and p = pinherent

Check feasibility

constraintsConverged?

noyesEstimate τ from throughput and p

Estimate p from throughput andτ

Output:feasible/infeasible

Input: throughput

13

Fair rate allocationFair rate allocationInitialization: add all demands to unsatSet

Scale up all demands in unsatSet until some demand is saturated or scale1

Output X

Move saturated demands from unsatSet to X

if (unsatSet≠)

if (scale 1)yes

no

yes

no

14

Total throughput maximizationTotal throughput maximization• Formulate a non-linear optimization problem

(NLP)

• Solve NLP using iterative linear programming

*0

2/)(1

10

)1(

)1(..

max

d

i

xx

pCW

TDT

pEPxRts

x

d

i

jjjijslot

jj

iii

ddid

dd

Sending rate is feasible

E2e throughput is bounded by demand

Link load is bounded bythroughput constraints

Maximize total throughput

15

Evaluation methodologyEvaluation methodology• Model validation

– How to quantify over-prediction error?• Verify if prescribed rates are achievable

– How to quantify under-prediction error?• Scale up all prescribed rates by a common factor

• Performance optimization– Fairness maximization: Jain’s fairness index– Total throughput maximization

• This talk: testbed results only– 19 mesh nodes at UTCS building; up to 7 hops– Extensive simulation results are in the paper

16

Optimization schemesOptimization schemes• Our rate optimization• No rate optimization (current practice)• Conflict graph based optimization

– Plug conflict graph model to our framework

– Conflict graph assumes perfect scheduling [JPPQ03]• Represent each wireless link with a vertex• Draw an edge between the vertices if the

corresponding links interfere• Derive clique constraints – all links in a clique

in the CG cannot be active together

17

Baseline: conflict graph modelBaseline: conflict graph model

0

2

4

6

8

10

12

1 2 3 4 5 6 7 8 9 10 11Actua

l throu

ghpu

t (Mbp

s)

Estimated throughput (Mbps)

0 2 4 6 8

10 12 14

0 2 4 6 8 10 12 14Actua

l throu

ghpu

t (Mbp

s)

Estimated throughput (Mbps)

CG model significantly over-estimates sending rates.

UDP TCP

y=0.8x

y=x y=x

y=0.8x

18

Model validation: UDP trafficModel validation: UDP traffic

0 1 2 3 4 5 6 7 8 9

10

0 1 2 3 4 5 6 7 8 9 10Actua

l throu

ghpu

t (Mbp

s)

Estimated throughput (Mbps)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1

0 0.2 0.4 0.6 0.8 1

Frac

tions

of r

unsRatios between actual and estimated throughput

scale=1.0scale=1.1scale=1.2scale=1.5

1) Most estimated rates are achievable within 20%.2) Rates scaled up by just 10% become unachievable.

y=x

y=0.8x

19

Model validation: TCP trafficModel validation: TCP traffic

0 1 2 3 4 5 6 7 8

0 1 2 3 4 5 6 7 8Actua

l throu

ghpu

t (Mbp

s)

Estimated throughput (Mbps)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1

0 0.2 0.4 0.6 0.8 1Fr

actio

ns of

runs

Ratios between actual and estimated throughput

scale=1.0scale=1.1scale=1.2scale=1.5

Our model is accurate for TCP traffic.

y=x

y=0.8x

20

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12 14 16Fa

irness

inde

x

Num of Flows

wo/ opt.CG opt.Our opt.

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12 14 16

Fairn

ess in

dex

Num of Flows

wo/ opt.CG opt.Our opt.

Maximizing fairnessMaximizing fairnessUDP TCP

Fairness index is close to 1 under our scheme, while it degrades quickly in other schemes.

21

0.5 1

1.5 2

2.5 3

3.5 4

4.5 5

5.5

0 2 4 6 8 10 12 14 16

Throu

ghpu

t

Num of Flows

wo/ opt.CG opt.Our opt.

Maximizing total throughputMaximizing total throughputUDP

Our scheme significantly increases total throughput.

1 1.5

2 2.5

3 3.5

4 4.5

5

0 2 4 6 8 10 12 14 16Th

rough

put

Num of Flows

wo/ opt.CG opt.Our opt.

TCP

22

ConclusionsConclusions• Main contributions

– Predictable wireless performance optimization• A simple yet accurate wireless network model• Effective model-driven optimization algorithms

– Demonstrate their effectiveness through testbed experiments and simulation

• Future work– Handle dynamic traffic and topologies– Use passive measurement to seed our

model

Thank you!Thank you!

24

0.5 1

1.5 2

2.5 3

3.5 4

4.5 5

5.5

0 2 4 6 8 10 12 14 16

Throu

ghpu

t (Mbp

s)

# Flows

wo/ opt.w/ our opt.

0.5 1

1.5 2

2.5 3

3.5 4

4.5 5

0 2 4 6 8 10 12 14 16Th

rough

put (M

bps)

# Flows

wo/ opt.w/ our opt.

Impact on different routing Impact on different routing schemesschemes

Our scheme helps all routing schemes considered.

TCPUDP

25

TCP Pathologies under no rate TCP Pathologies under no rate controlcontrol

S1

S2

R

D1

D2

No Rate Limit (Mbps)

Rate Limit

0.805, 0.740 1.066, 1.064

TCP cannot set the rate that maximizes throughput.