A Case Study of the Running Time Fluctuation of Application

Abstract

During the operation of a supercomputer, running time of an application occasionally becomes longer or shorter than that previously measured under the same conditions. This is called "running time fluctuation." Running time fluctuations disturb the efficient operation of a supercomputer and waste the consumption of precious computer resources. Thus, we have been investigating and resolving this issue on the K computer. These issues may occur for many applications and supercomputers. In this paper, we describe some causes of running time fluctuations that we encountered in the past.