SAREK: Optimistic Parallel Ordering in Byzantine Fault Tolerance

Recently proposed Byzantine fault-tolerant (BFT) systems achieve high throughput by processing requests in parallel. However, as their agreement protocols rely on a single leader and make big efforts to establish a global total order on all requests before execution, the performance and scalability of such approaches is limited. To address this problem we present SAREK, a parallel ordering framework that partitions the service state to exploit parallelism during both agreement as well as execution. SAREK utilizes request dependency which is abstracted from application-specific knowledge for the service state partitioning. Instead of having one leader at a time for the entire system, it uses one leader per partition and only establishes an order on requests accessing the same partition. SAREK supports operations that span multiple partitions and provides a deterministic mechanism to atomically process them. To address use cases in which there is not enough application-specific knowledge to always determine a priori which partition(s) a request will operate on, SAREK provides mechanisms to even handle mis-predictions without requiring rollbacks. Our evaluation of a key-value store shows an increase in throughput performance by a factor of 2 at half the latency compared to a single-leader implementation.