Revisiting SRPT for Job Scheduling in Computing Clusters

Abstract

As the scheduling principle of Shortest Remaining Processing Time (SRPT) has been proven to be optimal in the single-machine setting, it’s a natural thought that SRPT shall also be extended to yield various scheduling algorithms with theoretical performance guarantees in distributed computing clusters which consist of multiple machines. In this paper, we revisit the SRPT scheduling principle to derive new and tight competitive performance bounds with respect to the overall job flowtime. In particular, for the transient scheduling scenario where all jobs arrive at the cluster at time zero, we study two different cases and show that the SRPT-based scheduling algorithm can achieve a constant competitive ratio of at most two, compared to the prior state-of-the-art ratio of 12 in the algorithm of Moseley et al. For online scheduling, we study a special case where each job only consists of one single task and show that the online SRPT Algorithm is \((1+\epsilon )\)-speed, \((3 + \frac{3}{\epsilon })\)-competitive with respect to the overall job flowtime for \(\epsilon > 0\), improving the recent result of Fox and Moseley which upper bounds SRPT to be \((1+\epsilon )\)-speed, \(\frac{4}{\epsilon }\)-competitive.