Adaptive Content Delivery for Scalable Web Servers Authors: Rahul Pradhan and Mark Claypool Presented by: David Finkel Computer Science Department Worcester.

Adaptive Content Delivery for Scalable Web Servers

Authors: Rahul Pradhan and Mark Claypool

Presented by: David Finkel

Computer Science Department

Worcester Polytechnic Institute

Worcester, MA, USA

Outline

Introduction Managing LoadManaging Load Approach - Methodology - Design Experiments Client – Side Experiments Future Work Conclusions

Introduction

Enormous Growth of Web Exponential in number if users,

pages, and sites

New Web uses require different resources Static: Disk, Network Dynamic: CPU, Network Media streaming: CPU, Disk+,

Network+

Web Server Overload

Loaded server rejects connections, denying service Companies lose revenue

Loaded server increases response times Slower pages viewed as less

interesting

Current Approaches to Managing Load

Over provisioning Beef up single server Can still become loaded during flash crowds

Load balancing Server farm or CDN Individual servers may still become over-

loaded

Content adaptation Reduce resources needed upon heavy load

Content Adaptation Examples

Using “thumbnails” instead of full, inline imagesReducing the number of local links Reducing the number of embedded objectsChanging the quality of imagesCurrent approaches manual! Our approach is to adapt content

automatically

Our Approach

Two versions of Web pages High quality to serve under normal load Low quality to serve under high load

Monitor the server load A separate light weight standalone process

Upon heavy load, server switches to low quality transparentlyRequires no modification to Server Browser http protocol

Adaptive Content Delivery System Architecture

Load Monitor

Adaptation Module

Content Switching

CPU, Disk, Network

Web Server

Disk

Requests

Response

Base System Our System

Load Monitor

Continuously monitors the utilization of the server and the observed response timeDeveloped utilities to measure utilization CPU, Network, Disk Observed Response Time

Used Linux /proc file system, but techniques general enough for any OS

Adaptation Module

Input of load values from the load monitorDecides low or high load Low and high thresholds

Threshold values determined by prior measurement Scripts to induce load on server httperf to generate requests measured response time (using httperf)

Response Time vs CPU Util

Thresholds 60% and 75%

Content Selector TransparentlyTransparently switches contentswitches content depending on

the decision made by the adaptation module.

We use symbolicsymbolic linkslinks to make the same file point

to different qualities of content.

indexhigh.html

index.html indexlow.html

ExperimentsServer P-III, 500,128 MB RAM, IDE, 10 Mpbs, Linux

2.2.14, Apache 1.3.12

Workloads Static Workload Dynamic Workload Multimedia Workload

Metrics Throughput (Responses/sec) Average Response Time Percentage of Errors Frame Rate (for Multimedia Clients)

Response Time (ms) vs Requests/sec

Percent Errors vs Requests/sec

Frame rate vs Number of MM Clients

Overhead For the Adaptive Content Delivery System

Client – Side Experiments

Experiments on realreal servers to determine the impact of file size on response time.

Used a modified httperfhttperf for our measurements to generate requests

Measured the response time along with the connection set up time and transfer

time.

Conclusions

Server load criticalWe present a mechanism to Quantify server load Adapat transparently to client

Improves server performance: supports 25% more static requests supports twice as many Multimedia

clients supports 15% more CGI requests

Future Work

Adapting to heterogeneous client environment Clients may have different

bandwidths

Adding QoS features to the Web server Range of content quality at server Maximize QoS for user

Adaptive Content Delivery for Scalable Web Servers

Authors: Rahul Pradhan and Mark Claypool

Presented by: David Finkel

Computer Science Department

Worcester Polytechnic Institute

Worcester, MA, USA

Extra slides past here ….

JPEG Quality : JPEG Quality Factor vs Percentage Savings in File Size

MPEG Quality : MPEG Q Scale Factor vs Percentage File Size Savings

Response Time vs Number of CGI Requests

Percentage Of CGI Requests Rejected vs Number of CGI Requests