A heuristic for scheduling task graphs with communication delays onto multiprocessors

Abstract

The multiprocessor scheduling problem can be stated as finding a schedule for a task graph to be executed on a multiprocessor architecture so that the execution time can be minimized. Since this problem is known to be NP-hard, in all but a few very restricted cases, the main research efforts in this area are focused on heuristic methods for obtaining near-optimal solutions in a reasonable amount of time. A new compile-time single-pass multiprocessor scheduling technique, called chaining, has been developed and is presented in this paper. Chaining takes into account the communication overhead and can be applied to scheduling task graphs onto fully-connected multiprocessor architectures containing an arbitrary (bounded as well as an unbounded) number of processors. This technique can be viewed as a generalized list scheduling concept, that does not impose any preconditions about the ordering in which tasks are selected for scheduling as well as about the position within the current partial schedule where selected task can be placed. Varying the selection policy, implemented in this technique, we are able to generate a class of scheduling algorithms. As a representative example of this class we present Task Selection First (TSF) scheduler. We compare performances of the TSF scheduler with the dynamic level scheduler proposed by Sih and Lee, the dominant sequence clustering algorithm proposed by Yang and Gerasoulis, and the DSC/MLS algorithm, a modified version of Sarkar's two-step scheduling technique.