Living with nondeterminism in replicated middleware applications

Abstract

Application-level nondeterminism can lead to inconsistent state that defeats the purpose of replication as a fault-tolerance strategy. We present Midas, a new approach for living with nondeterminism in distributed, replicated, middleware applications. Midas exploits (i) the static program analysis of the application's source code prior to replica deployment and (ii) the online compensation of replica divergence even as replicas execute. We identify the sources of nondeterminism within the application, discriminate between actual and superficial nondeterminism, and track the propagation of actual nondeterminism. We evaluate our techniques for the active replication of servers using micro-benchmarks that contain various sources (multi-threading, system calls and propagation) of nondeterminism. © IFIP International Federation for Information Processing 2006.