We recently proposed a novel method for large-object replication and load balancing. Our method is particularly well-suited to data grids, data warehousing providers, and hosting of dynamic web sites. The method attempts to distribute object request load fairly to servers according to server capacity so that the likelihood of them overloading, and hence failing, is reduced. Unfortunately, server failures cannot be eliminated entirely. When a server fails, the load carried by that server must be absorbed by the rest of the system. Unless this load is distributed fairly across the remaining servers, they may also overload, creating a cascade of failures and reduced quality of service. In this paper, we propose an efficient method for fairly redistributing the load of a failed server or set of failed servers within our replication system. We also report on experimental results that verify the validity of our approach.
|Original language||English (US)|
|Number of pages||14|
|Journal||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|State||Published - Dec 1 2003|
ASJC Scopus subject areas
- Theoretical Computer Science
- Computer Science(all)