A Performance Study of Static Task Scheduling Heuristics on Cloud-Scale Acceleration Architecture
Yang Shi, Zhaoyun Chen, Wei Quan, Mei Wen · 2019
Accelerator based computing platforms have been frequently employed to meet the ever-increasing performance requirements of cloud computing. However, due to their complexity, the efficiency of these platforms is still an open research question. As a major solution for improving the efficiency of a computing platform, task scheduling, which has been studied extensively in recent decades, should be investigated in the context of accelerator-based cloud computing platforms. For this purpose, this paper compares 10 typical task scheduling algorithms on a classic cloud computing platform proposed by Microsoft, using both random and real application task sets. These 10 task scheduling algorithms can be divided into three categories: the Round Robin algorithm, list-based heuristics and computing-intensive searching algorithms. Experimental results show that searching algorithms outperform the other categories on small task sets; however when the scale of the scheduling problem increases towards that of real situations, list-based heuristics yield better results. Of these, Min-min and CROP are the most prominent.