A Case for Fluid Replication
Brian D. Noble, Ben Fleis, Minkyong Kim · 1999
The performance of distributed services is becoming increasingly variable due to changing load patterns and user mobility. The two approaches to this problem, cluster-based, scalable services and peer replication, solve only part of the problem. Cluster-based services deal only with end-service variability, while peer replication compromises safety and precludes the ability to offer bounds on consistency of updates. We propose fluid replication, automatic creation of replicas where and when they are needed, as a solution to this problem. In this paper, we present our mechanisms for finding replica sites, balancing consistency and performance, and maintaining client consistency when changing replicas. 1. Introduction As distributed services scale, their performance becomes increasingly unpredictable. There are three reasons for this. First, aggregate user interests tend to change rapidly and unpredictably, placing varying demands on services. Second, changing patterns of demand in the...