Performance characterization of decentralized algorithms for replica selection in distributed object systems

Abstract

Designers of distributed systems often rely on replicas for increased robustness, scalability, and performance. Replicated server architectures require some technique to select a target replica for each client transaction. In this paper, we use simulation to characterize performance (response time, selection error, probability of server overload) for four common replica-selection algorithms (random, greedy, partitioned, weighted) when applied in a decentralized form to client queries in a distributed object system deployed on a local network. We introduce two new selection algorithms (balanced and balanced-partitioned) that give improved performance over the more common algorithms. We find the weighted algorithm performs best among the common algorithms and the balanced algorithm performs best among all those we considered. Our findings should help designers of distributed object systems to make informed decisions when choosing among available replica-selection algorithms.