Performance sensitivity of space-sharing processor scheduling in distributed-memory multicomputers

Abstract

Processor scheduling in distributed-memory systems has received considerable attention in recent years. Several commercial distributed-memory systems use space-sharing processor scheduling. In space sharing, the set of processors in a system is partitioned and each partition is assigned for the exclusive use of a job. Space sharing policies can be divided into fixed, static, or dynamic categories. For distributed-memory systems, dynamic policies incur high overhead. Thus, static policies are considered as these policies provide better performance than the fixed policies. Several static policies have been proposed in the literature. In a previously proposed adaptive static policy, the partition size is a function of the number of queued jobs. This policy, however, tends to under-utilize the system resources. To improve the performance of this policy, we propose a new policy in which the partition size is a function of the total number of jobs in the system, as opposed to only the queued jobs. The results presented demonstrate that the new policy performs substantially better than the original policy for the various workload and system parameters. Another major contribution is the evaluation of the performance sensitivity to job structure, variances in inter-arrival times and job service times, and network topology.