Improving performance of parallel transaction processing systems by balancing data load on line

Abstract

The performance of parallel transaction processing systems can be degraded significantly due to data skew, a phenomenon of unbalanced data distribution over the nodes of a system. Rebalancing the data load of a system with data skew by redistributing its data is known to be an effective approach to cope with data skew. Unfortunately, for most of the existing approaches, the data being redistributed is unavailable (off-line). Numerous applications, such as those for reservations, finance, process control, hospitals, police and the armed forces, however, cannot afford off-line data for any significant amount of time. These applications call for the ability of balancing data load online. In this paper, a new online data redistribution approach is proposed. A prototype of the approach has been implemented, and experiments have been conducted. Experimental results confirm the substantial performance gains of the approach.