A Comparative Study of Multiprocessor List Scheduling Heuristics

This paper presents a quantitative comparison of a collection of DSP multiprocessor list scheduling heuristics which consider inter-processor communication delays. The following aspects have been addressed: (1) performance in terms of the total execution time (makespan), (2) sensitivity of heuristics in terms of the characteristics of acyclic precedence graphs, including graph size and graph parallelism, (3) sensitivity of heuristics to the number of processors, and (4) compile time e ciency. In addition, the e ectiveness of list scheduling performance enhancement techniques is examined. The main contributions of this paper are: Contrary to the belief of some previous authors, our study indicates that no single published list scheduling heuristic consistently produces the best schedules under all possible program structures and DSP multiprocessor congurations. We believe this fact is very important to designers of DSP multiprocessor scheduling heuristics. Based on such observations, we propose the DS scheduling method which, instead of using a single heuristic, allows a compiler to select di erent heuristics during the course of scheduling. Such selection is based on the number of executable tasks and available processors | quantities which change dynamically during the scheduling process. By selecting among simple heuristics, DS is also able to achieve a much faster compile time than other heuristics of comparable performance. Finally, to our knowledge, this paper provides a rst comprehensive simulation study comparing a wide range of list scheduling heuristics (including the DS method proposed in this paper) and several enhancement techniques for DSP multiprocessor architectures with non-zero inter-processor communication delay. We have selected 7 representative list scheduling heuristics including DS, as well as 3 performance enhancement techniques | yielding a total of 27 heuristic combinations each with 350 randomly generated acyclic precedence graphs. i