Abstract

Several data replication strategies have been proposed to provide high data availability for database system applications. However, the tradeoffs among the different strategies for various workloads and different operating modes is still not well understood. In this paper, we study the relative performance of three high availability data replication strategies, chained declustering, mirrored disks, and interleaved declustering, in a shared nothing database machine environment. Among the issues that we have examined are (1) the relative performance of different strategies when no failures have occurred, (2) the effect of a single node failure on system throughput and response time, (3) the performance impact of varying the CPU speed and/or disk page size on the different replication strategies, and (4) the tradeoff between the benefit of intra query parallelism and the overhead of activating and scheduling extra operator processes. Experimental results obtained from a simulation study indicates that, in the normal mode of operation, chained declustering and interleaved declustering perform comparably. Both perform better than mirrored disks if an application is I/O bound (due to disk scheduling), but slightly worse than mirrored disks if the application is CPU bound. In the event of a disk failure, because chained declustering is able to balance the workload while the other two cannot, it provides noticeably better performance than interleaved declustering and much better performance than mirrored disks. 1.

Citations