On GPU optimizations of stencil codes for highly parallel simulations

Stencil codes are valuable methods to solve partial differential equations of models in a wide range of applications in science and engineering. Graphics processing units (GPUs) provide a highly parallel architecture with fast directly accessible memory that is desirable to run stencil codes. They enable larger and more complex simulations that are solved faster compared to simulating on CPUs. We provide a solution for users to run highly parallel stencil codes on GPUs, that can also be efficiently used on a distributed GPU cluster.In this work, we present an extension to the multi-disciplinary framework NAStJA originally designed to efficiently run stencil codes on the CPUs in current high-performance computing (HPC) systems, to also be able to run on GPUs. We describe different methods to increase the performance of stencil codes on GPUs like a border exchange method which is transparent to the user, as well as a buffer for gradient values that are needed multiple times. We show their performance using the phase-field method as a stencil code example.With this GPU extension and optimizations implemented into the NAStJA framework, we can show highly improved performance compared to the CPU implementation, and an efficiency of 92% (weak scaling) for a large-scale example simulation on 64 GPUs on the ForHLR II HPC system.