Document Type

Technical Report

Publication Date


Technical Report Number



In this paper we examine the effects of replication on the availability of data in a large network. This analysis differs from previous analyses in that it compares the performance of a dynamic consistency control protocol not only to that of other consistency control protocols, but also to the performance of non-replication and to an upper bound on data availability. This analysis also differes in that we gather extensive simulations on large networks subject to partitions at realistically high component reliabilities. We examine the dynamic consistency protocol presented by Jajodia and Mutchler [9, 12] and by Long and Paris[18] along with two proposed enhancements to this protocol[10,11]. We study networks of 101 sites and up to 5050 links (fully-connected) in which all components, although highly reliable, are subject to failure. We demonstrate the importance in this realistic environment of an oft neglected parameter of the system model, the ratio of transaction submissions to component failures. We also show the impact of the number of copies on both the protocol performance and the potential of replicaion as measured by the upper bound. Our simulations show that the majority of current protocol performs optimally for topologies that yield availabilities of at least 65%. On the other hand, the availability provided by non-replicaion is inferior to that of the majority of current protocol by a most 5.9 percentage points for these same topologies. At this point of maximum difference, theprimary copy protocol yields availability 59.1% and the majority of current protocol yields availability 65.0%. We discuss the characteristics of the model limiting the performance of replication.