Accelerating Plasma Physics with GPUs

JuSPIC is a particle-in-cell (PIC) code, developed in the Simulation Lab for Plasma Physics of the Jülich Supercomputing Centre. The open source code is based on PSC by H. Ruhl, slimmed-down and rewritten in modern Fortran. JuSPIC simulates particles under the influence of electromagnetic fields, using the relativistic Vlasov equation and Maxwell's equations (integrated using the Finite Difference Time Domain scheme). The program uses a regular mesh for the Maxwell fields and the particle charge/current densities. Inside the mesh, quasi-particles with continuous coordinates are modeled via distribution functions. JuSPIC is part of the High-Q club, attesting that it can efficiently scale to the full JUQUEEN supercomputer (currently the #13 on the Top 500 list): 1.8 million threads running on 458 thousand cores can collaboratively compute plasma simulations. Local node-level parallelism is achieved by means of OpenMP, communication between nodes relies on MPI. To leverage the latest generation of supercomputers coming equipped with dedicated accelerator technologies (GPUs and other many-core architectures), JuSPIC is currently being extended. In this poster we present a GPU-accelerated version of the program, making use of different programming models. We show first results of performance studies, comparing OpenACC and CUDA. While OpenACC aims to offer portability and flexibility by means of few changes to the code, the performance of the generated program might suffer in practice. To measure the deficit, the compute-intensive parts of the program are in addition also implemented in CUDA Fortran. To explore scalability properties of the application for static particle distributions on a heterogeneous architecture, we make use of semi-empirical performance models