Towards Sustainable HPC: Energy Aware Benchmarking of FWI CPU Power Caps and Thread Trade-offs

This study presents an energy-aware benchmarking of 3D orthogonal Full Waveform Inversion (FWI) on the Shaheen III-CPU supercomputer, targeting sustainable high-performance computing (HPC) practices. FWI is a computationally intensive seismic imaging method, and its scalability demands careful evaluation under power-constrained environments. The benchmarking involves 28 configurations, combining seven power caps (200W–400W) and four OpenMP thread counts (48–192). Performance was assessed by measuring execution time and energy usage (in watt-hours), revealing key trade-offs between speed and energy efficiency.

Results demonstrate that increasing thread count improves performance only under sufficient power caps (≥280W). Beyond 144 threads, however, gains diminish or reverse, especially at lower power caps due to parallel efficiency limitations. Notably, the configuration with maximum resources (192 threads at 400W) did not yield the best performance-per-watt. Instead, mid-range configurations (e.g., 144 threads at 280–320W) provided an optimal balance between runtime and energy consumption.

The findings underscore that for energy-constrained scientific computing, maximizing hardware usage does not guarantee best performance. Instead, tuning for power efficiency is essential. These insights serve as a guide for selecting sustainable configurations for seismic imaging and other compute-intensive applications on modern HPC systems.