N-Body simulation using GP-GPU: evaluating host/device memory transference overhead

Abstract

N-Body simulation algorithms are amongst the most commonly used within the field of scientific computing. Especially in computational astrophysics, they are used to simulate gravitational scenarios for solar systems or galactic collisions. Parallel versions of such N-Body algorithms have been extensively designed and optimized for multicore and distributed computing schemes. However, N-Body algorithms are still a novelty in the field of GPGPU computing. Although several N-body algorithms have been proved to harness the potential of a modern GPU processor, there are additional complexities that this architecture presents that could be analyzed for possible optimizations. In this article, we introduce the problem of host to device (GPU) – and vice versa – data transferring overhead and analyze a way to estimate its impact in the performance of simulations.