Performance and energy task migration model for heterogeneous clusters

Abstract

This article presents a set of linear regression models to predict the impact of task migration on different objectives, like performance and energy consumption. It allows to establish whether at a given moment the migration of a task is profitable in terms of performance or energy consumption. Also, it can be used to determine the best node to migrate a task depending on the objective. The model uses a small set of parameters that are easily measurable. It has been validated against a small heterogeneous cluster using the Slurm resource manager. The model captures the tendencies observed in the results of the experiments, with average relative errors below 3.5% in execution time and 2.5% in energy consumption.