Classification of HPC Job Power Consumption Using a Rich Feature Set

High-performance computing (HPC) systems consume significant amounts of power, and accurately predicting the power consumption of individual jobs can be a valuable tool for reducing energy costs. While many efforts have focused on power prediction using common features from standard HPC job traces, there remains a lack of integration of the rich, contextual information available exclusively in job scripts. In this work, we present an alternative approach by training a model to classify HPC jobs as either high- or low-power consumers using a feature set enriched by job script data. By extracting information such as primary software used, loaded modules, and activated Conda environments directly from job scripts, we incorporate features not previously leveraged for this task. This enables more robust and reliable classification, especially as job heterogeneity and complexity increase in modern HPC systems. We demonstrate the efficacy of this method on a large dataset of 1.5 million jobs run on a petascale cluster over a nine-month period, achieving an overall accuracy of 89% and a median accuracy of 97% for individual models trained and tested over time. We also release our complete feature extraction and modeling codebase as an open resource, enabling reproducibility and collaborative advancement in energy-aware HPC operations.