On the optimality of degree of declustering

Abstract

We present a data partitioning technique for shared-nothing database systems. A unique feature of our scheme is that it organizes a multicomputer system into groups of even number of PNs; and each relation is assigned to one of these groups in such a way to minimize contention among concurrent queries. Thus, a fixed degree of declustering is used for all base relations in this scheme. Our simulation results demonstrate that this approach provides significantly better performance than those of conventional methods which independently determine a degree of declustering for each of the base relations. These schemes totally ignore the requirement of interquery parallelism. Obviously, an appropriate degree of declustering represents a good trade-off between interquery and intraquery parallelism for our strategy. To investigate this issue, we perform extensive simulations to study the effect of various system and workload parameters on the optimality of the degree of declustering. We found that it is influenced primarily by the parallel processing overhead. With this finding, we develop a mathematical model to determine the optimal degree of declustering for a given system.