An efficient predicting delay algorithm for short jobs

Abstract

Traditional delay scheduling algorithm processes the task scheduling through a fixed time waiting threshold to achieve better data locality. However, this fixed manner often leads to low resource utilization and job performance degradation, especially for short jobs. This paper proposes a predicting delay scheduling algorithm, called P-Delay. P-Delay leverages both the job characteristic and dynamic resource availability to accelerate the short job execution. It provides a dynamic delay scheduling by comparing the transferring cost with the predicted waiting cost to make reasonable wait. Experimental results show that P-Delay can achieve a speedup of 1.45× for short jobs than traditional Delay Scheduling algorithm.