Dynamic Scalable State Machine Replication

Abstract

State machine replication (SMR) is a well-known technique that guarantees strong consistency (i.e., linearizability) to online services. In SMR, client commands are executed in the same order on all server replicas: after executing each command, every replica reaches the same state. However, SMR lacks scalability: every replica executes all commands, so adding servers does not increase the maximum throughput. Scalable SMR (S-SMR) addresses this problem by partitioning the service state, allowing commands to execute only in some replicas, providing scalability while still ensuring linearizability. One problem is that ssmr quickly saturates when executing multi-partition commands, as partitions must communicate. Dynamic S-SMR (DS-SMR) solves this issue by repartitioning the state dynamically, based on the workload. Variables that are usually accessed together are moved to the same partition, which significantly improves scalability. We evaluate the performance of DS-SMR with a scalable social network application.