Investigating the use of GPUs with a Monte Carlo Astrophysical Simulation
University of Edinburgh
University of Edinburgh, 2013
@article{willis2013investigating,
title={Investigating the use of GPUs with a Monte Carlo Astrophysical Simulation},
author={Willis, James S},
year={2013}
}
For a given simulation, the most expensive subroutine in the astrophysics code, MOCCA (MOnte Carlo Cluster SimulAtor), has been ported to run as a kernel on a GPU (Graphics Processing Unit). The code was accelerated using the CUDA programming model, which was performed with PGI CUDA Fortran. The GPU code was run with varying problem sizes and a number of optimisations were performed. It was successfully run on a GPU cluster that contained NVIDIA’s Tesla C2050 GPU cards. The best GPU version of the do loop took 1.84x longer than the serial version for the smaller problem size and the code ran 1.01x faster for the larger problem size. Suggestions have been made for further optimisations, which could be implemented to achieve an improvement in performance.
November 25, 2013 by hgpu
