Workshops Archive

Predicting the Energy-Optimal GPU Clock for Production VASP Jobs Before They Run


Workshop: Sustainable Supercomputing

Authors: Kevin Menear (National Laboratory of the Rockies)

Abstract: VASP is a dominant HPC workload, and while many users are migrating their workflows to GPUs, not every job uses the hardware efficiently. Lowering the GPU clock for these jobs can reduce their energy consumption with limited impact on performance. However, the right clock differs from job to job. We present a model for per-job energy optimization, one that chooses each job’s clock from its input files alone, before the job runs. Every job’s slowdown follows the same machine-wide curve as the clock drops, scaled by a per-job sensitivity, and this shared curve carries 99.72% of the variance across jobs. Hence choosing the clock needs only the sensitivity, predicted from the submitted files. Under a 5% slowdown budget the selection saves 14.25% of GPU-board energy on 310 production GH200 jobs at 5.45% mean slowdown, and the median job lands on budget at 4.92%. The error concentrates in the most clock-sensitive jobs, whose sensitivity is underpredicted, but it falls as the corpus grows, so the remedy is more training data. Finally, the model transfers. We show sensitivity carries across machines, so a single calibration job maps the model onto an A100, and a handful of jobs settles the delivered clocks.


Back to Sustainable Supercomputing Archive Listing

Back to Full Workshop Archive Listing