Manetho: Fault Tolerance in Distributed Systems Using Rollback-Recovery and Process Replication

Abstract

This dissertation presents a new protocol that allows rollback-recovery and process replication to co-exist in a distributed system. The protocol relies on a novel data structure called the antecedence graph, which tracks the nondeterministic events during failure-free operation and provides information for recreating them if a failure occurs. The rollback-recovery part of the protocol combines the low failure-free overhead of optimistic rollback-recovery with the advantages of pessimistic rollback-recovery, namely fast output commit, limited rollback, and failure-containment. The process replication part of the protocol features anew multicast protocol designed specifically to support process replication. Unlike previous work, the new protocol provides high throughput and low latency in message delivery without relying on the application semantics. The protocol has been implemented in the Manetho prototype. Experience with a number of long-running, compute-intensive parallel applications confirms the performance advantages of the new protocol. The implementation also features several performance optimizations that are applicable to other rollback-recovery and multicast protocols.