Predicting Job Power Consumption Based on RJMS Submission Data in HPC Systems
Résumé
Power-aware scheduling is a promising solution to the resource usage monitoring of High-Performance Computing facility electrical power consumption. This kind of solution needs a reliable estimation of job power consumption to feed the Resources and Jobs Management System at submission time. Available data for inference is restricted in practice because unavailable or even untrustworthy. We propose in this work an instance-based model using only the submission logs and user provided job data. GID and the number of tasks per node appears to be good features for prediction of a job’s average power consumption. Moreover, we extant this model to production context with online computation to make a practical global power prediction from job submission data using instances re-weighting. The performance of the online model are excellent on COBALT’s data. With any doubt this model will be a good candidate for the achievement of consistent power-aware scheduling for other computing centers with similar informative inputs.