SSR: Synchronous State Machine Replication with a Timing-Aligned Split Architecture

Abstract

Modern datacenters exhibit different timing properties across the network fabric and the host software stack: the former offers bounded-delay delivery and precise clock synchronization, whereas the latter provides greater flexibility but substantially greater timing variability. This substrate-level separation matches the functional structure of State Machine Replication (SMR), whose normal case is timing-sensitive and repetitive while recovery requires flexible coordination, state transfer, and membership reconfiguration. We present SSR, a synchronous state machine replication protocol. SSR exploits this alignment with a split protocol design: normal-case replication runs as a round-based synchronous SMR protocol close to the network fabric, where replicas propose concurrently and commit from per-round local evidence without a steady-state leader, while recovery remains in a host-side path. Our preliminary evaluation results show that SSR improves common-case throughput by 14.9× with three replicas and 24.9× with five replicas, completes recovery in about 0.15 ms, and commits on the fast path after two rounds.