State-Machine and Deferred-Update Replication: Analysis and Comparison

Abstract

In the paper, we analyze and experimentally compare two popular replication schemes relying on atomic broadcast: state machine replication (SMR) and deferred update replication (DUR). We estimate the lower bounds on the time of executing requests by the SMR and DUR systems running on multi-core servers. We also consider variants of systems that can process read-only requests with a lower overhead. In the analysis of DUR, we consider conflict patterns. We then formally show the scalability of SMR and DUR, which reflects the capacity of systems to effectively utilize an increasing number of processor cores. Next, we compare SMR and DUR experimentally under different levels of contention, using several benchmarks. We show throughput, abort rate (in DUR), and network congestion. The key results of our work are that neither system is superior in all cases, and that the theoretical and experimental results are heavily influenced by the dominance of either the CPU execution time or atomic broadcast time. We therefore propose to combine both replication schemes and gain the best of both worlds.