high performance computing on graphics processing units: hgpu.org

The potential for decreasing the solution time for the UK Met Office NAME III [1] lagrangian particle atmospheric particle dispersion modelling code was examined. The code was ported to the EPCC Ness and Fermi0 machines and compiled with the PGI compiler. Timing benchmarks and profiling completed for a particle only run, and a cloud gamma run to examine potential areas for speed up. A prototypical simple dispersion model was conceptually compared to the NAME III Particle benchmark. This simple model was accelerated using OpenMP, CUDAC and CUDA FORTRAN. Timing benchmarks and profiling completed for various problem sizes from 1000, to 10,000,000 particles, 10,000,000 representing a realistic problem size for an emergency particle run [2]. The simple model was found to have a total speed up of up to ~50x for the largest problem size with the particle loop being sped up ~80-100x. The results can be extrapolated to indicate the NAME III code could be sped up by approximately 12x for this specific benchmark, or about 59x scaled to a realistic problem size. Other areas with potential for speed up in NAME III were also evaluated. Due to the complexity of the NAME III code although CUDA and GP-GPU acceleration can readily be applied to targeted areas of code representing specific benchmarks, it may not represent the best option for speeding up the whole of the code. The Met Office may like to consider a hybrid OpenMP and MPI approach utilising the existing OpenMP implementation.