GPU-Accelerated MPI-Parallelized Grid-Based Vlasov Simulations of Collisionless Plasmas in Electric Propulsion Modeling: Preliminary Progress

Grid-based Vlasov simulations provide a high-fidelity, the collisionless plasmas in EP systems, but their practical application is limited by extensive computational expense. This paper details our preliminary progress in developing a GPU-accelerated, MPI-parallelized version of our in-house code, Vlasolver, to overcome this challenge. To ensure performance portability across different GPU architectures from vendors like NVIDIA, AMD, and Intel, we adopt the Kokkos C++ programming model. This approach required a significant redesign of key computational kernels to enable massive parallelism and avoid data races. Preliminary unoptimized benchmarks of a 2D plasma beam expansion problem show that the GPU-accelerated code on eight GPUs achieves around 10∼15x speedup over the MPI-parallelized CPU code running on 800 cores. Initial unoptimized strong scaling results for the 2D version of EP plume expansion are also presented and show comparable strong-scaling trends. This work demonstrates the potential of GPU acceleration to make the Vlasov method a more viable tool for predictive EP modeling.