Lay Summary

Computational Fluid Dynamics (CFD) applications are notoriously demanding, relying heavily on raw CPU performance and fast memory access. This paper evaluates the performance of modern AMD EPYC processor architectures when executing complex fluid dynamics software. By conducting extensive benchmarks across various mesh sizes and parallel computing setups, the researchers analyze processor throughput, memory bandwidth usage, and scalability. Their analysis offers practical guidance on configuring supercomputing systems and software parameters to achieve optimal execution speeds for engineering and environmental fluid simulations. (82 words)

Full Abstract

Memory bandwidth and cache hierarchy design strongly dictate the execution efficiency of memory-bound Computational Fluid Dynamics (CFD) codes on multi-core microprocessors. This paper evaluates recent AMD EPYC CPU architectures (including 3D V-Cache variants) using representative HPC fluid dynamics solvers. We perform detailed benchmark experiments across varying mesh densities, core counts, and memory configurations. The benchmark results illustrate how memory bandwidth saturation impacts parallel scaling and highlight hardware tuning settings—such as NUMA configuration and memory channel binding—necessary to achieve maximum throughput for complex CFD problems.

Metadata

• Publication Date: 2024
• Author: Marcin Lawenda, Łukasz Szustak, László Környei, Flavio Cesar Cunha Galeazzo, Paweł Bratek
• Journal: Performance Evaluation / Supercomputing Benchmarks