Resource Utilization Analysis of Early Scheduling in Parallel State Machine Replication

Most approaches to scaling the throughput of state machine replication resort to concurrent execution of non-conflicting requests. While these approaches enhance throughput, conflict detection and handling introduce overhead. Early scheduling is a promising technique that trades concurrency for expeditious scheduling decisions. In this technique, requests are grouped in classes and a fixed subset of threads is assigned to each class, respecting request dependencies. Early scheduling has been shown to provide significant performance improvements in state machine replication. This paper evaluates the impact of the restrictions imposed by the early scheduling technique. In particular, it shows that threads may be idle while pending independent requests are available to be executed, leading to poor processor utilization. We characterize resource underutilization for workloads with different rates of conflicting requests, number of threads, and number of request classes. The paper opens up new opportunities to further enhance the early scheduling technique.