Gang scheduling performance on a cluster of non-dedicated workstations

Clusters of workstations have emerged as a cost-effective solution to high performance computing problem. To take advantage of any opportunities, however, effective scheduling techniques are necessary that enable parallel applications to dynamically share workstations with their owners. In this paper a special type of parallel applications called gangs are considered. Gangs are jobs that consist of a number of interacting tasks scheduled to run simultaneously on separate and distinct processors. A simulation model is used to address performance issues associated with gang scheduling on a non-dedicated workstation cluster for various workloads. Simulated results indicate that the relative performance of the gang scheduling policies that we examine depends on the workload characteristics.