Top Banner
Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000
12

Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

Jan 06, 2018

Download

Documents

Chester Bridges

3 Introduction Alexandria digital library Geographically-referenced materials Map, satellite image, digitized aerial photograph Large data Require extensive processing Goal : minimize response time
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

Adaptive Load Sharing for Clustered Digital Library Servers

Song, yong-joo

System Software Lab, EECS, KAIST9. 13, 2000

Page 2: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

Contents Introduction Previous Work

Clustered ADL server design Load sharing strategy Policy for node selection & load collection

Experiment Conclusion & Critique

Page 3: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

3

Introduction Alexandria digital library

Geographically-referenced materials Map, satellite image, digitized aerial photograph

Large data Require extensive processing

Goal : minimize response time

Page 4: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

4

Introduction(cont’d)

Basic ADL Clustered ADL

Page 5: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

5

Page 6: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

6

Previous Work Clustered Web server design

DNS round-robin rotation Problem : load imbalance

SWEB Redirecting request by dynamic system load prediction of response time

Page 7: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

7

SWEB DNS round-robin rotation Redirect request to minimum expected

response time server

CPU, I/O load monitoring

prediction of DB searching time load information update

overhead

Page 8: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

8

Load Sharing strategy Class of request

Text & image ( CPU & I/O usage small ) DB search ( CPU & I/O usage maybe large ) Image processing ( predictable )

Relative Server-site Response Cost Availability relative to the unloaded base node

redirection threshold value

atioDiskAvailRdthDiskBandwindwidthBaseDiskBa

IdleRatioCPUSpeededBaseCPUSpeRSRC

)1(

Page 9: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

9

Policy for load collection Broadcast – redundant info, load explosion Random poll – synchronization overhead

Proposed Random multicast

A hybrid approach Random multicast + Random poll

)1)(1()1( atioDiskAvailRIdleRatio

Page 10: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

10

Experiment Server

6 of 147Mhz 2 of dual 167Mhz

Multi-node performance Comparison of load collection policies Effectiveness of RSRC func.

Page 11: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

11

Conclusion Study adaptive load balancing strategy

Consider both I/O & CPU load factor On non-uniform cluster Hybrid node selection & collection policy

Page 12: Adaptive Load Sharing for Clustered Digital Library Servers Song, yong-joo System Software Lab, EECS, KAIST 9. 13, 2000.

12

Critique SWEB does not consider non-uniform

cluster? RSRC correctness Parameter, threshold value decision Multicast overhead

How about Broadcast & selective catch