On the Effect of Server On the Effect of Server Adaptation Adaptation

for Web Content Deliveryfor Web Content Delivery

IMW ’02, Marseille, Nov. 2002IMW ’02, Marseille, Nov. 2002

Joint work withJoint work withBalachander Krishnamurthy (AT&T)Balachander Krishnamurthy (AT&T)

Craig Wills (WPI)Craig Wills (WPI)

Yin Zhang (AT&T)Yin Zhang (AT&T)

11/8/2002 IMW 2002 2

MotivationMotivation

Web sites have a strong incentive to reduce time-to-glass

Challenge client connectivity is

heterogeneous Natural solution –

server adaptation client connectivity +

content characteristics + client capability + server load + … action to take

InternetInternet

foobar.comfoobar.com

11/8/2002 IMW 2002 3

Study: What?Study: What? Basic question – What exactly is the

performance impact of server adaptation? When and how much can server adaptation help? Which action should the server take?

Lots of previous work … but typically focusing on one individual action

This study – Provides a unified framework for assessing the

impact of different server actions Obtains useful insights through multi-site wide-

area measurements

11/8/2002 IMW 2002 4

Factors ConsideredFactors Considered Client connectivity

Latency, bandwidth Content characteristics

Criteria: total bytes, container bytes, #objects 3x3x3 = 27 buckets

derived from large proxy logs further justified by examining popular Websites’ pages

Server actions Altering the content

reducing number of images, reducing image size Altering the location of the content

using a Content Distribution Network (CDN) Altering manner of delivery

compression, bundling Altering protocol options

using persistent connections Combination of different actions

11/8/2002 IMW 2002 5

Experiment MethodologyExperiment Methodology A multi-site study

Server: Apache West coast: icir East coast: wpi

Client: httperf US: att, modem, isdn Intl: de, au, uk

Canonical content served at each site

covering the space of buckets

Experiments repeated at different times of day0

50

100

150

200

250

300

350

0 20 40 60 80 100 120 140

Roun

d-Tr

ip T

ime (

ms)

Throughput (KB/sec)

att.icir

att.wpi(462.6)

de.icir

de.wpiisdn.icir

isdn.wpi

modem.icir

modem.wpi

uk.icir

uk.wpi

au.icir

au.wpi

11/8/2002 IMW 2002 6

ResultsResults Compression of HTML is not universally useful

It only works for bandwidth-constrained clients Persistent connections alone has limited benefit

Little improvement for all client/server combo Pipelining gives significant improvement

Exception: bandwidth-constrained clients Bundling gives significant improvement

Bundling alone is similar to pipelining Compressed bundles help a lot under all conditions CDN-served bundles – good idea for well-connected

clients Reducing image size by half has little benefit Reducing the number of objects by half helps a lot

under most conditionsBaseline: 4 parallel HTTP/1.0 connectionsBaseline: 4 parallel HTTP/1.0 connections

11/8/2002 IMW 2002 7

Contribution and Further Contribution and Further WorkWork

Contribution A unified framework for evaluating the

impact of server adaptation Can be applied by individual Web site

Insights we gained can be useful for improving client performance

Further work Evaluation of the feasibility of online client

classification and server adaptation through real implementation

Our results are encouraging

11/8/2002 IMW 2002 8

AcknowledgmentsAcknowledgments People who gave us accounts / logs