Lay Summary

Running massive Computational Fluid Dynamics (CFD) simulations on modern supercomputers consumes vast amounts of electrical energy. To make high-performance computing more sustainable, scientists must balance execution speed with energy consumption. This paper introduces a predictive model designed to analyze the trade-off between performance and energy efficiency for CFD software running on AMD-based computer clusters. By forecasting how different hardware configurations and parallelization settings impact energy draw, the model helps researchers select optimal operational parameters that accelerate simulation times without incurring excessive power costs. (84 words)

Full Abstract

Energy consumption has become a limiting factor in high-performance computing (HPC), requiring energy-aware optimization alongside traditional runtime scaling. In this paper, we construct a performance–energy trade-off prediction model tailored for Computational Fluid Dynamics (CFD) applications executed on AMD EPYC-based clusters. Using statistical profiling and machine learning regression, the model estimates runtime duration and energy footprint based on mesh resolution, core allocation, and CPU frequency scaling. The predictive framework enables system operators and scientists to identify energy-optimal deployment strategies that minimize power consumption while meeting strict execution deadlines.

Metadata

• Publication Date: 2024
• Author: Marcin Lawenda, Łukasz Szustak, László Környei
• Journal: Sustainable Computing: Informatics and Systems / HPC Proceedings