Harvesting energy consumption on European HPC systems: Sharing Experience from the CEEC project

TL;DR

This paper tackles the energy footprint of HPC by presenting CEEC’s practical framework for measuring and interpreting energy use across major European systems. It combines hardware counters, software tools, and two core metrics—Energy Ratio and EPID—to enable reproducible, energy-aware benchmarking, demonstrated through CFD-focused case studies on waLBerla, FLEXI, Neko/Nekbone, and NekRS across LUMI, MareNostrum5, MeluXina, and JUWELS Booster. Key findings show accelerators and mixed-precision techniques can substantially reduce energy-to-solution while maintaining accuracy, though efficiency can decline with underutilization or extreme device counts, highlighting the need for balanced workloads and standardized measurement practices. Collectively, the work advocates for accessible, standardized energy measurements to guide sustainable exascale computing and informs hardware/software design decisions with real-world CFD workloads.

Abstract

Energy efficiency has emerged as a central challenge for modern high-performance computing (HPC) systems, where escalating computational demands and architectural complexity have led to significant energy footprints. This paper presents the collective experience of the EuroHPC JU Center of Excellence in Exascale CFD (CEEC) in measuring, analyzing, and optimizing energy consumption across major European HPC systems. We briefly review key methodologies and tools for energy measurement as well as define metrics for reporting results. Through case studies using representative CFD applications (waLBerla, FLEXI/GALÆXI, Neko, and NekRS), we evaluate energy-to-solution and time-to-solution metrics on diverse architectures, including CPU- and GPU-based partitions of LUMI, MareNostrum5, MeluXina, and JUWELS Booster. Our results highlight the advantages of accelerators and mixed-precision techniques for reducing energy consumption while maintaining computational accuracy. Finally, we advocate the need to facilitate energy measurements on HPC systems in order to raise awareness, teach the community, and take actions toward more sustainable exascale computing.

Figures (11)

• Figure 1: Hierarchical relation between energy measurement methods.
• Figure 2: Three-dimensional numerical setup of LHC4, waLBerlakemmlerTowardsExascaleSimulations2026.
• Figure 3: Time-to-solution (left) and energy-to-solution (right) on a single node run of LHC4, waLBerla.
• Figure 4: Time-to-solution (left) and energy-to-solution (right) on a single node run for the dominating simulation modules of LHC4, waLBerla.
• Figure 5: Energy efficiency (EPID) and parallel efficiency (PID) of FLEXI/GA-LÆ-XI on the CPU (left) and GPU (right) partition of MeluXina.