Octo-Tiger: Binary star systems with HPX on Nvidia P100

Abstract

Stellar mergers between two suns are a significant field of study since they can lead to astrophysical phenomena such as type Ia supernovae. Octo-Tiger simulates merging stars by computing self-gravitating astrophysical fluids. By relying on the high-level library HPX for parallelization and Vc for vectorization, Octo-Tiger combines high performance with ease of development. For accurate simulations, Octo-Tiger requires massive computational resources. To improve hardware utilization, we introduce a stencil-based approach for computing the gravitational field using the fast multipole method. This approach was tailored for machines with wide vector units like Intel's Knights Landing or modern GPUs. Our implementation targets AVX512 enabled processors and is backward compatible with older vector extensions (AVX2, AVX, SSE). We further extended our approach to make use of available NVIDIA GPUs as coprocessors. We developed a tasking system that processes critical compute kernels on the GPU or the processor, depending on their utilization. Using the stencil-based fast multipole method, we gain a consistent speedup on all platforms, over the classical interaction-list-based implementation. On an Intel Xeon Phi 7210, we achieve a speedup of 1.9x. On a heterogeneous node with an Intel Xeon E5-2690 v3, we can obtain a speedup of 1.46x by adding an NVIDIA P100 GPU.