Performance analysis of DEM simulations on modern supercomputers: a case study

Large-scale Discrete Element simulations are highly compute-intensive, and achieving energy-efficient performance remains a major challenge for next-generation exascale systems. This work evaluates the performance and energy characteristics of DEM kernels on multinode systems using MESA-PD within waLBerla and P4IRS, a portable framework optimized for CPUs and GPUs. While MESA-PD efficiently supports CPU simulations, it lacks GPU portability. Integrating P4IRS into waLBerla enables GPU-accelerated and portable DEM execution. Using a settling-spheres benchmark, we study performance and energy consumption across different systems, achieving up to 90% weak-scaling efficiency on 16 CPU nodes and demonstrating scalability to 256 GPUs.