Eecient Local Scheduling of Parallel Tasks and Communications on Heterogeneous Networks of Workstations

The network of workstations (NOW) we consider for local scheduling of parallel tasks and communications is heterogeneous and nondedicated, where computing power varies among the workstations, and multiple jobs may interact with each other in execution. Local scheduling is a standard method on networks of workstations, where each node in the system makes independent scheduling decisions. Coordinating parallel tasks and minimizing communication delays are two important and diicult issues for local scheduling. We address the two issues by presenting a scheduling technique, called self-coordinated local scheduling. To coordinate parallel tasks, each local scheduler schedules both parallel tasks and sequential user jobs based on a static power preservation in each workstation. Thus, the parallel tasks in one computational phase can be guaranteed to complete within a time period. To minimize the communication latency, each local scheduler uses a non-preemptive strategy | tries to make one communication activity complete in a continuous time-frame. Our experiments show that the power preservation in each workstation improves the performance of both sequential and parallel jobs, and network utilization. Furthermore, the non-preemptive communication strategy reduces general communication contention and delay.