Fairly redistributing failed server load in a distributed system

Abstract

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. © Springer-Verlag Berlin Heidelberg 2003.