Performance of a Class of Highly-Parallel Divide-and-Conquer Algorithms

Abstract

A wide range of important problems have efficient methods of solution based on the divide-and-conquer strategy. However, the traditional approach to parallel divide-and-conquer does not scale well due to the sequential component of the algorithms. We investigate the performance of a non-traditional, highly-parallel divide-and-conquer strategy that we refer to as "one-deep parallel divide-and-conquer". As representative examples, we measure and analyze the performance of highly-parallel variants of the quicksort and mergesort algorithms on a shared-memory multiprocessor. Both algorithms deliver impressive speedups. We present a systematic approach to modeling the performance of one-deep parallel divide-and-conquer algorithms and we demonstrate this approach with the two parallel sorting algorithms.