Grid services often consist of nondeterministic operations. The traditional way to replicate nondeterministic services is using the primary-backup approach that requires the system be synchronous, which is often not true for grid environments. This paper addresses the problem of replicating nondeterministic services on asynchronous systems. We design a protocol based on Paxos and propose two approaches to optimize the service performance for the requests that do not change service states and the requests using transactions. The evaluation on a local cluster at UCSD shows that both of these optimizations significantly improved the request latency and service throughput of replicated services.
Xianan Zhang - PhD Student, UCSD
Flavio Junqueira- PhD Student, UCSD
Matti Hiltunen - Researcher, AT&T
Keith Marzullo - Professor, UCSD
Rick Schlichting - Researcher, AT&T
last modified: 01/11/2006