Greifswald 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
P: Fachverband Plasmaphysik
P 1: Magnetic Confinement I/HEPP I
P 1.2: Talk
Monday, February 26, 2024, 11:30–11:55, ELP 6: HS 3
GPU offloading strategies for gyrokinetic edge turbulence simulations with GENE-X via OpenMP and OpenACC — •Jordy Trilaksono1, Carl-Martin Pfeiler1, Philipp Ulbl1, and Frank Jenko1,2 — 1Max Planck Institute for Plasma Physics, Boltzmannstraße 2, 85748 Garching, Germany — 2University of Texas at Austin, Austin, TX 78712, USA
The GENE-X code simulates plasma turbulence by solving the gyrokinetic equation using a grid-based/Eulerian discretization. The flux-coordinate independent approach allows GENE-X to simulate plasma turbulence anywhere within magnetic confinement fusion (MCF) devices, from the plasma core to the wall. GENE-X is mainly written in object-oriented modern Fortran 2008 fully utilizing MPI+OpenMP parallelization. Here, we present our development strategies and experiences to further accelerate GENE-X on GPU, which is essential for simulations towards larger, reactor-relevant fusion devices. The GPU offloading features are written on an auxiliary C++ layer interoperated by the main Fortran layer. The C++ layer provides broader selections of GPU offloading tools. MPI+OpenMP and MPI+OpenACC parallelizations are chosen to future-proof our solution against the evolution and diversification of modern GPU architectures. We present performance benchmarks and convergence analysis of our OpenMP and OpenACC implementations on GPU. The computational hotspot in GENE-X achieves a significant performance increase on GPU compared to its CPU-equivalent. The readiness of GENE-X compute capability for large-scale production runs on GPU is further investigated.
Keywords: Gyrokinetic; Plasma edge physics; High Performance Computing; GPU computing; Heterogeneous memory