Structured partitioning of concurrent programs for execution on multiprocessors

Abstract

In this paper, the problem of partitioning of parallel programs for execution on multiprocessors is investigated. By assuming run-time scheduling approaches, the static partitioning problems are formulated and solved in the context of a structured program representation model, in which a program is assumed to be composed of parallel loops. The structured partitioning problem is then defined as the problem of partitioning each parallel loop into appropriate number of tasks such that the program execution time is minimized in some sense. The worst case of the program execution cost is adopted as the minimization criterion, so the results obtained in this paper are guaranteed to be within some performance bound. Two algorithms are developed to solve this problem. The first algorithm using a prune-and-search technique can find the optimal partition, while the second algorithm can obtain a near optimal partition for the simplified partition problem in linear time. It is proved that the cost of the partition generated by the linear time algorithm is at most twice the cost of the optimal partition.