Plasma-PEPSC D2.5 Updated report on Phase-1 and initial results for Phase-2 codesign on EPI systems

This deliverable, D2.5, reports on the codesign status of plasma simulation codes (BIT1, Vlasiator, PIConGPU, GENE, GENE-X) on EPI systems, targeting both EPI Processors and Accelerators. For BIT1, optimising data locality on EPI Processors showed performance gains up to 19%. On the EPI Accelerator, manual vectorization of key kernels led to significant overall application speedups of 86% to 92% on a 2-node setup, depending on the precision-performance trade-off chosen for reductions. Vlasiator developments for EPI Processors included supporting Arm SVE via VectorclassSVE, achieving a speedup of 41\% for compute-intensive benchmarks. EPI Accelerator work with RVV 1.0 specification suggested a substantial reduction in instruction count. For PIConGPU, we focused on the underlying parallelization library alpaka and its sucessor alpaka3. We clearly identified the vectorization challenges introduced by alpaka's CUDA-inspired programming model and how they are overcome in its successor alpaka3. On PIConGPU level, we identified how the design trade-offs currently favor many-core environments over vector-centric hardware and discussed how recent "enabled-by-exascale" algorithmic additions to PIConGPU are likely to shift this towards a more vectorization-friendly design in the future. We added alpaka3's ability to dynamically determine RISC-V vector length for performance. GENE scaling on EPI Processors showed efficient scaling only along specific dimensions (W, Z, V), with poor overall scalability. On the EPI Accelerator, analysis revealed that the main bottleneck region is not vectorized, and vectorized parts use low vector lengths, necessitating code restructuring. Initial work on GENE-X for the EPI Accelerator found that autovectorization severely degraded performance due to the main bottleneck operator having very short inner loops, resulting in ineffective, low-vector-length vectorization and numerical instability. Rewriting this operator is required.