Parallelism and concurrency control performance in distributed database machines

Abstract

While several distributed (or 'shared nothing') database machines exist in the form of prototypes or commercial products, and a number of distributed concurrency control algorithms are available, the effect of parallelism on concurrency control performance has received little attention. This paper examines the interplay between parallelism and transaction performance in a distributed database machine context. Four alternative concurrency control algorithms are considered, including two-phase locking, wound-wait, basic timestamp ordering, and optimistic concurrency control. Issues addressed include how performance scales as a function of machine size and the degree to which partitioning the database for intra-transaction parallelism improves performance for the different algorithms. We examine performance from several perspectives, including response time, throughput, and speedup, and we do so over a fairly wide range of system loads. We also examine the performance impact of certain important overhead factors (e.g., communication and process initiation costs) on the four alternative concurrency control algorithms.