Plasma-PEPSC D1.9 First Release – Plasma Simulation Codes and Documentation

In this deliverable, we publicly release current versions of the codes within the project and summarize the contributions to the codes made since the beginning of Plasma-PEPSC two years ago. In short, BIT1 development was split into four parts: optimization (vectorization) of a number of functions, implementation of the parallel I/O via openPMD-API and ADIOS-2 libraries, load balancing and GPU porting. The first two tasks were (almost) completed, while the others are under development. The development work for GENE and GENE-X was mainly put into the GPU porting of the GENE-X code, which will take most of the envisaged runtime of the target run. The code was fully ported with an OpenMP directives approach, and now the time loop completely runs on NVIDIA GPUs. The extension to AMD GPUs (like in LUMI-G) is in development. The development of PIConGPU is focused on realizing the exascale science case, with proposals for LUMI and JUPITER currently under consideration. The project includes the implementation of a high-precision laser model and associated solvers, which are expected to reach completion by 2025. Full-scale simulation runs have already been planned to take advantage of these advances. To enhance the performance of PIConGPU, significant updates have been made to its parallelism infrastructure, which is based on the alpaka library, including improvements to vectorization support. The on-device memory allocation scheme has been updated using MallocMC. To address the I/O and analysis bottleneck, work was done to enable PIConGPU, integrated with openPMD, to function as part of an in situ machine learning (ML) workflow, and this was successfully demonstrated in a full-scale run. PIConGPU has also already demonstrated its scalability and capability through full-scale runs on LUMI this year. Vlasiator's development focussed on performance portability onto both vector-heavy and GPU-heavy target platforms, with a major restructuring of solver data structures and flow control behavior to be both flexible and future-proof. In parallel, algorithmic improvements of both the Vlasov and field solvers substantially improved the HPC scaling behavior of the code. New, more dynamic adaptive mesh refinement methods allow the simulations' resolution to track areas of physical interest more effectively. New in situ data reducers for magnetic topology identification and non-Maxwellianity measures allow for reduced output file footprints. The adoption of more structured development practices and more end-user outreach has greatly improved documentation and usability.