Optimizing Cost and Performance for Multihoming ACM SIGCOMM 2004 Lili Qiu Microsoft Research [email protected] Joint Work with D. K. Goldenberg, H. Xie,

Optimizing Cost and Optimizing Cost and Performance for Performance for

MultihomingMultihoming

ACM SIGCOMM 2004ACM SIGCOMM 2004

Lili QiuLili QiuMicrosoft ResearchMicrosoft [email protected]@microsoft.com

Joint Work withJoint Work withD. K. Goldenberg, H. Xie, Y. R. Yang, Yale D. K. Goldenberg, H. Xie, Y. R. Yang, Yale

University University Y. Zhang, AT&T Labs – ResearchY. Zhang, AT&T Labs – Research

2

Multihoming & Smart RoutingMultihoming & Smart Routing

Multihoming– A popular way of

connecting to Internet

Smart routing– Intelligently distribute

traffic among multiple external links

User

ISP 1

ISP K

InternetISP 2

3

Potential BenefitsPotential Benefits• Improve performance

– Potential improvement: 25+% [Akella03]– Similar to overlay routing [Akella04]

• Improve reliability– Two orders of magnitude improvement in

fault tolerance of end-to-end paths [Akella04]

• Reduce cost

Q: How to realize the potential benefits?

4

Our GoalsOur Goals• Goal

– Design effective smart routing algorithms to realize the potential benefits of multihoming

• Questions– How to assign traffic to multiple ISPs to

optimize cost? – How to assign traffic to multiple ISPs to

optimize both cost and performance?– What are the global effects of smart

routing?

5

Related WorkRelated WorkTechniques for implementing multihoming

– BGP peering, DNS-based, NAT-based (e.g., [RFC2260, Cisco, GCLC04, Radware, F5])

– Complementary to our workPerformance evaluation [Akella03,Akella04]

– Quantify the potential benefits of multihoming– Unaddressed challenge: how to achieve this in

practiceSmart routing

– Commercial products (e.g., [RouteScience, Internap, Proficient, …])

– Technical details are unavailableHash-based load balancing [Cao01, Guo04]

– Optimizes neither performance nor cost

6

Network ModelNetwork Model• Network performance metric

– Latency (also an indicator for reliability)– Extend to alternative metrics

• log (1/(1-lossRate)), or latency+w*log(1/(1-lossRate))

• ISP charging models– Cost = C0 + C(x)– C0: a fixed subscription cost– C: a piece-wise linear non-decreasing function

mapping x to cost – x: charging volume

• Total volume based charging• Percentile-based charging (95-th percentile)

7

Percentile Based ChargingPercentile Based Charging

Interval

Sorted volume

N95%*N

Charging volume: traffic in the (95%*N)-th sorted interval

8

Why cost optimization?Why cost optimization?• A simple example:

– A user subscribes to 4 ISPs, whose latency is uniformly distributed

– In every interval, the user generates one unit of traffic

• To optimize performance– ISP 1: 1, 0, 0, 0, …– ISP 2: 0, 1, 0, 0, …– ISP 3: 0, 0, 1, 0, …– ISP 4: 0, 0, 0, 1, …– 95th-percentile = 1 for all 4 ISPs– 95th-percentile = 1 using one ISP

• Cost(4 ISPs) = 4 * cost(1 ISP)

Optimizing performance alone could result in high cost!

9

Cost Optimization: Cost Optimization: Problem Specification (2 ISPs)Problem Specification (2 ISPs)

Time

Volume

N1 2

10

Cost Optimization: Cost Optimization: Problem Specification (2 ISPs)Problem Specification (2 ISPs)

Time

Volume

P1

P2

Goal: minimize total cost = C1(P1)+C2(P2)

Sorted volume

Sorted volume

11

Issues & InsightsIssues & Insights• Challenge: traditional optimization techniques

do not work with percentiles

• Key: determine each ISP’s charging volume

• Results– Let V0 denote the sum of all ISPs’ charging volume – Theorem 1: Minimize cost Minimize V0

– Theorem 2: V0 ≥ 1- k=1..N(1-qk) quantile of original traffic, where qk is ISP k’s charging percentile

12

Cost Optimization: Cost Optimization: Problem Specification (2 ISPs)Problem Specification (2 ISPs)

Time

Volume

P1

P2P1 + P2 90-th percentile of original traffic

Sorted volume

Sorted volume

13

Intuition for 2-ISP CaseIntuition for 2-ISP Case• ISP 1 has 5% intervals whose traffic exceeds

P1• ISP 2 has 5% intervals whose traffic exceeds

P2

• The original traffic (ISP 1 + ISP 2 traffic) has 10% intervals whose traffic exceeds P1+P2

• P1+P2 90-th percentile of original traffic

14

Sketch of Our AlgorithmSketch of Our Algorithm1. Determine charging volume for each ISP

– Compute V0

– Find pk that minimize ∑k ck(pk) subject to ∑kpk=V0 using dynamic programming

2. Assign traffic given charging volumes– Non-peak assignment: ISP k is assigned pk

– Peak assignment: • First let every ISP k serve its charging volume pk

• Dump all the remaining traffic to an ISP k that has bursted for fewer than (1-qk)*N intervals

15

Additional IssuesAdditional Issues• Deal with capacity constraints

• Perform integral assignment– Similar to bin packing (greedy heuristic)

• Make it online– Traffic prediction

• Exponential weighted moving average (EWMA)– Accommodate prediction errors

• Update V0 conservatively• Add margins when computing charging

volumes

16

Optimizing Cost + PerformanceOptimizing Cost + Performance• One possible approach: design a metric

that is a weighted sum of cost and performance– How to determine relative weights?

• Our approach: optimize performance under cost constraints– Use cost optimization to derive upper

bounds of traffic that can be assigned to each ISP

– Assign traffic to optimize performance subject to the upper bounds

17

Evaluation MethodologyEvaluation Methodology• Traffic traces (Oct. 2003 – Jan. 2004)

– Abilene traces (NetFlow data on Internet2)• RedHat, NASA/GSFC, NOAA Silver Springs Lab,

NSF, National Library of Medicine• Univ. of Wisconsin, Univ. of Oregon, UCLA, MIT

– MSNBC Web access logs

• Realistic cost functions [Feb. 2002 Blind RFP]

• Delay traces– NLANR traces: 3 months’ RTT measurements

between pairs of 140 universities– Map delay traces to hosts in traffic traces

18

Baseline AlgorithmsBaseline Algorithms1. Round robin

– In each interval, assign traffic to a single ISP– Rotate in a round robin fashion

2. Equal split– In each interval, split traffic equally among ISPs– Similar to hash-based load balancing

3. Offline local fractional– Minimize the total cost for each interval

independently

4. Dedicated links– Flat rate and independent of usage

19

Cost Comparison for Different TracesCost Comparison for Different Traces

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

Red Hat MIT UCLA Wisconsin Web Server

No

rmal

ized

co

st t

offline cost online cost round robin equal LFA offline

Our algorithms significantly out-perform the alternatives.

20

Cost Comparison for Varying # Links Cost Comparison for Varying # Links

For all # ISPs, our cost optimization performs well.

0

1

2

3

4

5

6

7

8

0 5 10 15

# external links

No

rmalize

d c

ostl

offline cost online cost round robin

equal LFA offline

21

Cost + Performance EvaluationCost + Performance Evaluation

Optimizing performance alone often doubles the cost.

0

0.5

1

1.5

2

2.5

3

Co

st n

orm

aliz

ed b

y o

ffli

ne

cost

offline cost offline cost+perf online cost

online cost+perf offline perf

22

Cost + Performance Evaluation Cost + Performance Evaluation (Cont.)(Cont.)

Our dual metric optimization achieves low cost and latency.

0

10

20

30

40

50

60

70

Avera

ge L

ate

ncy (

ms)

offline cost offline cost+perf online cost online cost+perf offline perf

23

Global Effects of Smart RoutingGlobal Effects of Smart Routing• Selfish nature of smart routing

– Each user optimizes its own cost & performance without considering its impact on other traffic

– Need to understand its global effects

• Questions– How well does smart routing perform when traffic

assignment affects link latency?– How well do different smart routing users co-exist?– How well do smart routing users co-exist with

single-homed users?

24

Evaluation MethodologyEvaluation Methodology• Abilene traffic traces• Rocketfuel inter-domain topology

– 170 nodes, 600 edges– With propagation delay and OSPF weights – M/M/1 queuing model

• Routing– A user selects best performing ISP subject to

cost constraints– Inter-domain: shortest AS hop count– Intra-domain: OSPF

• Compute traffic equilibria as in [QYZS03]

25

Global Effects: SummaryGlobal Effects: Summary• Impact of self interference is small

• Smart routing users co-exist well with each other

• Smart routing users co-exist well with single-homed users

26

ConclusionsConclusionsContributions

– First paper on jointly optimizing cost and performance for multihoming

– Propose a series of novel smart routing algorithms that achieve both low cost and good performance

– Under traffic equilibria, smart routing improves performance without hurting other traffic

Future work– Further evaluation through Internet experiments– Dynamics of interactions among different users– Design better charging models

27

Thank you!