A new look at timestamp ordering concurrency control

Abstract

Popular conception has been that timestamp ordering concurrency control (CC) yields poor transaction processing performance in relation to twophase locking (2PL). This paper makes two contributions. First, we show the surprising result that Basic Timestamp Ordering (BTO) performs better than 2PL in all cases except when both data contention and message latency are low. When latency or data contention is high, BTO significantly outperforms 2PL. Our second contribution is a new timestamp ordering CC technique (PREDICT) that performs better than 2PL and BTO under low data contention, and continues to perform well under high data contention. We evolve a set of variants of PREDICT, and demonstrate that PREDICT achieves a good balance between lost opportunity cost and restart cost, and outperforms popular CC techniques.