Adaptive scheduling of computations and communications on distributed-memory systems

Compile-time scheduling is one approach to extract parallelism which has proved effective when the execution behavior is predictable. Unfortunately, the performance of most priority-based scheduling algorithms is computation dependent. Scheduling based on the concept of earliest-startable-task produces reasonably short schedules only when available parallelism is large enough to cover the communications. A priority-based decision is more effective when parallelism is low. We propose a scheduling in which the decision function combines two concepts: (1) task-level as global priority and (2) earliest-task-first as local priority The degree of dominance of one of the above concepts is controlled by a computation profile factor that is the ratio of task parallelism to communication. It is shown that the above factor is an upper bound on the deviation of schedule length from optimum. To tune the solution finish time the above scheduler is iteratively applied on the computation graph. In each iteration, the newly generated schedule is used to sharpen the task-levels which contribute in finding shorter schedules in the next iteration. Evaluation is carried out for a wide category of computation graphs with communications for which optimum schedules are known. It is found that pure local scheduling and static priority-based scheduling significantly deviate from the optimum under specific problem instances. Our approach to adapting the scheduling decision to the computation profile is able to produce near-optimum solutions via a much reduced number of iterations than other approaches.