Speculative client execution in deferred update replication

Abstract

Deferred Update Replication (DUR) is a powerful replication technique that allows parallelism of clients' execution while a global certification phase checks the validity of the transactional execution against workloads running on remote nodes. The well-known favorable scenario of DUR is when remote transactions rarely conflict with each other. In this paper we show that, even in this case, the conflicts happening among local application threads can significantly decrease performance. We address this problem by using speculation. We let local transactions propagate their post-execution snapshot to other local transactions before the outcome of the global certification is notified. This way, in scenarios where accesses are partitioned across nodes, we prevent local transactions from aborting each other. Through experimental study based on well-known transactional benchmarks we assess the effectiveness of the approach, gaining more than 10x using TPC-C benchmark.