Performance Evaluation of Scheduling Precedence-Constained Computations on Message-Passing Systems

Scheduling precedence graphs with communication times is the theoretical basis for achieving efficient parallelism in message-passing machines. The lack of global information on the tasks, due to communication, has lead to develop local scheduling heuristics such as the Earliest-Task-First. Using knowledge on computation, communication, and system topology, a class of global priority-based scheduling heuristics called Generalized List Scheduling is proposed. The task-level is evaluated by backward scheduling the computation over the multiprocessor by using the best local heuristic. This leads to realistic measurement of the task priority for use in forward GLS scheduling. Experimental evaluation of local and GLS heuristics is carried out using extensive random graph generation and altering over the communication, inherent parallelism, and system topology. Analysis shows that local heuristics rely on locally maximizing the processor efficiency and gives acceptable deviations only when the inherent parallelism is large enough to cover the effective communication. This leads the local heuristics to achieve bounded speedup. GLS scheduling is based on combining two strategies: 1) differentiate critical computation and communications from others by scheduling critical paths first, and 2) implement effective management of processor utilization in order to increase the speedup. GLS scheduling maintains acceptable relative deviation versus change in parallelism, communication, and multiprocessor topology. The time complexity of GLS heuristics is O(pn) , where p and n are the number of processors and that of the tasks, respectively.