A Performance Analysis of Large Scale Scientific Computing Applications from Log Archives

Abstract

A log archive for scientific computing applications is a set of logs for model and time of jobs in HPCs. We have developed light weight and fast performance analysis tools on top of log archives. We classify the job logs based on the similarity of the input models to form a model-based tree like archive. With linear regression, we analyze the relations of the step time of the jobs with the parameters in the model. We found that although there is some disturbance, the performance of most of the jobs showed good regularity. In one of the applications, we found the step time of job changes proportionally to the geometric parameters of model. And the most significant physical parameter determines step time up to 1.7 times. In another application, we find that the performance of each step scales 1.59 times with the number of process scales from 384 to 768.