Parallel State Machine Replication from Generalized Consensus

Abstract

State machine replication (SMR) is a established approach to building fault-tolerant services. In search for high SMR throughput, approaches that exploit semantic information in the ordering and execution of commands have emerged. Generalized consensus and parallel state machine replication are two representative examples, respectively. Although both approaches have been proved effective in isolation, no study in the literature has considered their integration. In this paper, we investigate the integration of generalized consensus and parallel SMR. We derive algorithms to parallelize the execution of commands based on the ordering of commands provided by consensus. As a prototype, we extended Egalitarian Paxos and conducted many experiments varying conflict rates, command computational costs, and number of cores at replicas. Compared to Egalitarian Paxos, the integrated approach (a) results in important throughput gains, as command independency and computational cost increase, and (b) converges to the same performance with high conflict rates or reduced number of cores. Index Terms– State Machine Replication, Generalized Consensus, Distributed Algorithms