Including accurate user estimates in HPC schedulers: ban empirical analysis

Abstract

This article focuses on the problem of dealing with low accuracy of job runtime estimates provided by users of high performance computing systems. The main goal of the study is to evaluate the benefits on the system utilization of providing accurate estimations, in order to motivate users to make an effort to provide better estimates. We propose the Penalty Scheduling Policy for including information about user estimates. The experimental evaluation is performed over realistic workload and scenarios, and validated by the use of a job scheduler simulator. We simulated different static and dynamic scenarios, which emulate diverse user behavior regarding the estimation of jobs runtime. Results demonstrate that the accuracy of users runtime estimates influences the waiting time of jobs. Under our proposed policy, in a scenario where users improve their estimates, waiting time of users with high accuracy can be up to 2.43 times lower than users with the lowest accuracy.