The Impact of Program Structure on the Performance of Scheduling Policies in Multiprocessor Systems ‡

A simple fork and join type of job structure has been extensively used for performance evaluation of processor scheduling policies in multiprocessor systems. However, parallel programs often exhibit a more complicated structure. It is not clear how the program structure affects the performance of processor scheduling policies. This paper studies the impact of the program structure on the performance of processor scheduling policies that are appropriate for uniform memory access (UMA) shared-memory systems. We consider four types of parallel program structures that are frequently employed in parallel applications. These are the fork-and-join, divide-and-conquer, Gaussian elimination, and state space search programs. The impact of these four job structures on the performance of four processor scheduling policies is considered. These are the job-based round robin (RRJob), coscheduling, dynamic (Dynamic), and the preemptive smallest cumulative demand first (PSCDF) policies. These policies cover a variety of policies − preemptive vs. non-preemptive, space-sharing vs. timesharing, a priori job knowledge vs. no job knowledge. The PSCDF policy is not implementable but is included for comparison purposes only. The results presented here show that the time-sharing processor scheduling policy RRJob performs as well as or better than the space-sharing Dynamic policy. This result contradicts the results of some previous studies and we clearly identify the problems associated with these previous studies. Thus, the major contribution of this paper is to correct the prevailing intuition that dynamic space-sharing outperforms time-sharing and show that, for the workloads considered here, time-sharing policies are useful for multiprocessor systems.