Performance studies of shared-nothing parallel transaction processing systems

We study the system performance affected by Degree of Declustering (DD) and the additional overheads in shared-nothing parallel transaction processing systems with the widely used two-phased locking (2PL) concurrency control method by simulation. We show that for the environment of high resource contention, a large DD is beneficial. For a system in the environment of low resource contention, if we have infinite frontend resource, a large DD is beneficial; otherwise a large DD would degrade system performance. It is shown that as the level of the resource contention in disks decreases, the additional overhead of parallelism increases. Furthermore, it is shown that the functions attached to frontend nodes affect system performance to a significant degree for shared-nothing parallel TP systems. These results provide insights for improving performance of shared-nothing parallel transaction processing systems.