The First OpenFOAM HPC Challenge (OHC-1)

TL;DR

The paper reports on the first OpenFOAM HPC Challenge, benchmarking computational performance across diverse hardware and software configurations, revealing key insights into energy efficiency and hardware/software optimization impacts.

Contribution

It introduces a large-scale benchmarking challenge for OpenFOAM, comparing hardware and software optimizations on diverse HPC systems with detailed performance analysis.

Findings

Hardware track identified Pareto-optimal performance-energy solutions.

On-package high-bandwidth memory (HBM) significantly improves single-node performance.

Software optimizations achieved up to 28% lower energy and 72% faster iteration times.

Abstract

The first OpenFOAM HPC Challenge (OHC-1) was organised by the OpenFOAM HPC Technical Committee (HPCTC) to collect a snapshot of OpenFOAM's computational performance on contemporary production hardware and to compare hardware-constrained submissions with software-track optimisations. Participants ran a common incompressible steady-state RANS case, the open-closed cooling DrivAer (occDrivAer) configuration, on prescribed meshes, submitting either with the reference setup (hardware track) or with modified solvers, decomposition strategies, or accelerator offloading (software track). In total, 237 valid datapoints were submitted by 12 contributors: 175 in the hardware track and 62 in the software track. The hardware track covered 25 distinct CPU models across AMD, Intel, and ARM families, with runs spanning from single-node configurations up to 256 nodes (32768 CPU cores). Wall-clock times ranged from 7.8 minutes to 65.7 hours and reported energy-to-solution from 2.1 to 236.9 kWh. Analysis of the hardware track identified a Pareto front of optimal balance between time- and energy-to-solution, and revealed that on-package high-bandwidth memory (HBM) dominates single-node performance for next-generation CPUs. Software-track submissions achieved up to 28% lower energy per iteration, 17% higher maximum performance per node, and 72% shorter minimum time per iteration than the best hardware-track results, with full GPU ports and selective-memory optimisations leading the performance range. This manuscript describes the challenge organisation, the case setup and metrics, and presents the main findings from both tracks together with an outlook for future challenges.