Design Tradeoffs of Alternative Query Tree Representations for Multiprocessor Database Machines

Abstract

During the past five years the design, implementation, and evaluation of joint algorithms that exploit large main memories and parallel processors has received a great deal of attention. However, the methods used to represent join queries and their corresponding effects on performance has received little attention during this same time span. In this paper we examine the tradeoffs imposed by left-deep, right-deep and gushy query trees in a multiprocessor environment. Specifically, we address potential parallelism, memory consumption, support for dataflow processing, and the cost of optimization that are dictated by a particular query tree format. Results indicate that for hash-based join algorithms, right-deep query trees provide the best potential to exploit large multiprocessor database machines.