Parallel Job Scheduling in Homogeneous Distributed Systems

The performance of parallel programs such as those involving fork-join instructions is significantly affected by the method used to schedule program tasks. This paper studies parallel job scheduling in homogeneous distributed systems. A simulation model is used to address performance issues associated with scheduling. Various policies are employed to schedule parallel jobs over a variety of workloads. Fairness is required among competing jobs. We examine cases where the distribution of the number of parallel tasks per job, and also the distribution of task service demand, vary with time. We also examine the impact of overhead necessary to collect global system information about processor queues on performance. Simulated results indicate that although all scheduling methods have merit, one significantly improves the overall performance and guarantees fairness in terms of individual job execution.