User estimates inaccuracy study in HPC scheduler

Abstract

The article is focused on user estimates inaccuracy study in high performance computing scheduler. The main goal of this survey is to study historical job submission data from the real supercomputer cenre and investigate accuracy of resources estimated by user on job submission. Based on this study 68.94% of jobs used less than 25% of their requested walltime, which indicates a high estimation inaccuracy. In case of estimation of amount of memory only 31.46% of jobs uses more than 75% of requested memory. Analysis of historical jobs from the further prediction of resources accuracy has been made. Similarities of the new submitted jobs with historical data were described by keys based on user names, software name, walltime and number of processors. Also mathematical model for jobs scheduling in HPC system, with a goal functions, has been proposed and verified.