Communication Scheduling in Parallel Task Executions on Large Parallel Systems

Scheduling is an important issue in parallel processing. Most scheduling algorithms makes assigns tasks in a direct acyclic graph (DAG) to processors. Usually only the allocation and ordering of tasks are considered, and sometimes communication time is included in the determination of priorities of the tasks, however, communication messages are not explicitly scheduled. Moreover, communication contention plays an increasing important role because of the increased system size of parallel computers. We study the effect of communication scheduling on large interconnection network and try to avoid communication contention. We present a communication scheduling scheme which can be easily included in exiting list-based scheduling algorithms. Experiments show that parallel execution time is reduced as compared to traditional scheduling without explicit communication scheduling.