Improving Processor Selection in List Scheduling Algorithms

This paper presents a technique to be applied to list scheduling algorithms which significantly reduces scheduling costs without sacrificing performance. In list scheduling algorithms, at each scheduling step one task is scheduled to the processor on which the task achieves the earliest start time. As a result, each task is tentatively scheduled to each processor to find the destination processor. We prove that only two tentative task scheduling are necessarily for each task to find the earliest task start time. As a consequence, the list scheduling complexity is significantly reduced from O(V log (W ) + EP ) to O(V (log (W )+log (P ))+E), where V andE are the number of tasks and edges in the task graph, respectively, W is the task graph width and P is the number of processors. We demonstrate through theory and experiments that the modified version of the scheduling algorithms using the described technique yields at a significantly lower cost equal or better performance compared to the original versions of the list scheduling algorithms.