DynaStar: Optimized Dynamic Partitioning for Scalable State Machine Replication

Classic state machine replication (SMR) does not scale well, since each replica must execute every command. To address this problem, several systems have investigated the use of state partitioning in the context of SMR, allowing client commands to be executed on a subset of replicas. Prior approaches range from completely static schemes, which do not adapt as workloads change, to dynamic schemes, which move data on-demand. This paper presents DynaStar, a new dynamic partitioning scheme for scaling state machine replication. In contrast to prior dynamic schemes, DynaStar uses a replicated location oracle to maintain a global view of the workload and inform heuristics about data placement. Using this oracle, DynaStar is able to adapt to workload changes over time, while also minimizing the number of state moves. The result is a practical technique that achieves excellent performance.