List Scheduling for Heterogeneous Computing Systems Introducing a Performance-Effective Definition for Critical Path
Farzam Dorostkar, Sattar Mirzakuchaki · 2019
The emergence of heterogeneous computing systems has been accompanied by serious design issues. Being a highly influential factor on performance in these systems, application scheduling is one of the major design considerations. In this paper, we propose a new critical path-oriented list scheduling heuristic algorithm called Communication-Intensive Path on a Processor (CIPOP) for heterogeneous computing environments. It is a modification of the well-known CPOP algorithm that presented the idea of scheduling the most costly entry-exit path of tasks, commonly known as the critical path, on a single processor. Generally, this processor selection strategy has different potential impacts on computation and communication costs along a selected path in the produced schedule. However, these probably different effects are not considered in the common definition of a critical path. Differentiating between these two types of costs, the proposed algorithm introduces a novel performance-effective definition for a critical path that is compatible with this processor selection strategy. CIPOP has a time complexity the same as that of the state-of-the-art list scheduling heuristic algorithms, which is of the order O(v2.× p) for v tasks and p processors. The conducted comprehensive experiment based on a wide variety of randomly generated application DAGs demonstrates the performance improvement of the proposed algorithm.