Predicting Job Power Consumption Based on RJMS Submission Data in HPC Systems
AI/Machine Learning/Deep Learning
Energy Consumption and Modeling
TimeMonday, June 22nd5:30pm - 6:00pm
DescriptionPower-aware scheduling is a promising solution to the increasingly problematic monitoring of High-Performance Computing (HPC) facility electrical power consumption. In this approach, the Resources and Jobs Management System (RJMS) requires a power consumption estimation for jobs at submission time. Available data for inference is restricted in practice to submission logs and user provided job data which is potentially unreliable. An instance-based model using only the reliable data is proposed, which can be used to demonstrate that the GID and the number of tasks per node are good features for prediction of a job's average power consumption. This model can easily be improved to make a global power prediction from job submission data using re-weighting of instances and then be adapted to the practical use-cases with online computations. The resulting model has a prediction error close to the measured value of XXX data. If results can be reproduced for other computing centers, this model will be a good candidate for the achievement of consistent power-aware scheduling.