This paper presents dispatching and scheduling algorithms for both periodic and aperiodic jobs on a cluster of computers. A cluster system consists of one front-end node, called dispatcher and multiple back-end nodes, called schedulers. By rigorous theoretical analysis, these algorithms are optimal in the sense that each aperiodic job can achieve the shortest response time in the cluster while maintaining the schedulability of periodic jobs. The aperiodic jobs, which have no deadline requirement, are partitioned into phases and assigned deadlines by the dispatcher. With the deadlines the aperiodic jobs can be scheduled uniformly under EDF together with periodic jobs.

Based on extensive simulation studies, we have that the algorithms proposed in this paper can achieve great performance improvement. We also work out the workload zone in which the response time of aperiodic jobs in the cluster with N schedulers is the same as the ideal bound of response time in the corresponding M/M/N model.