A list scheduling anomaly

Abstract

In the design of multiprocessor systems, the scheduling of tasks of a precedence-constrained task graph (representing a parallel program) onto the processors has a critical impact on overall system performance. The multiprocessor scheduling problem is known to be NP-hard in all but a few restricted cases. In this paper, we consider list scheduling heuristic algorithms for the multiprocessor scheduling problem. We show that these algorithms, while scheduling certain task graphs with non-negligible intertask communication, exhibit an anomaly which we call the list scheduling anomaly where the completion time of a task graph on a multiprocessor is more than its completion time on a single processor. Finally, we experimentally characterize the class of task graphs for which the list scheduling anomaly is more prominent.