Method of performing a parallel relational database query in a

Abstract

A method of performing a parallel join operation on a pair of relations R1 and R2 in a system containing P processors organized into Q clusters of P/Q processors each. The system contains disk storage for each cluster, shared by the processors of that cluster, together with a shared intermediate memory (SIM) accessible by all processors. The relations R1 and R2 to be joined are first sorted on the join column. The underlying domain of the join column is then partitioned into P ranges of equal size. Each range is further divided into M subranges of progressively decreasing size to create MP tasks T.sub.m,p, the subranges of a given range being so sized relative to one another that the estimated completion time for task T.sub.m,p is a predetermined fraction that of task T.sub.m-1,p. Tasks T.sub.m,p with larger time estimates are assigned (and the corresponding tuples shipped) to the cluster to which processor p belongs, while tasks with smaller time estimates are assigned to the SIM, which is regarded as a universal cluster (cluster 0). The actual task-to-processor assignments are determined dynamically during the join phase in accordance with the dynamic longest processing time first (DLPT) algorithm. Each processor within a cluster picks its next task at any given decision point to be the one with the largest time estimate which is owned by that cluster or by cluster 0.