Strengthening Atomic Multicast for Partitioned State Machine Replication

Partitioned state machine replication is a technique that extends classical state machine replication with state partitioning (or sharding) to provide both fault tolerance and performance scalability. The crux of the technique is ordering client requests within a partition, among the replicas that implement the partition, and across partitions, involving all the replicas accessed by the request. To cope with the complexity of ordering requests, partitioned state machine replication can use atomic multicast, a communication abstraction. Atomic multicast provides the means for requests to be propagated reliably and consistently to one or more sets of groups of replicas, where each replica group implements one partition. The paper revisits atomic multicast from the perspective of partitioned state machine replication and makes the following contributions: First, we show that if one implements partitioned state machine replication using an atomic multicast with global total order, a strong order property, then replicas would need to further coordinate as part of the execution of requests to ensure correctness. Second, we introduce a stronger version of atomic multicast that accounts for real-time dependencies between requests. Our proposed atomic multicast can be used to order requests within and across partitions so that replicas do not need to further coordinate to ensure linearizability. Third, we extend a well-known implementation of atomic multicast to ensure the stronger order property.