Accelerating Geophysics Simulation using CUDA
Brandon Holt, Daniel J. Ernst · The Journal of Computational Science Education · 2011
CitcomS, a finite element code that models convection in the Earth's mantle, is used by many computational geophysicists to study the Earth's interior.In order to allow faster experiments and greater simulation capability, there is a push to increase the performance of the code to allow more computations to complete in the same amount of time.To accomplish this we leverage the massively parallel capabilities of graphics processors (GPUs), specifically those using NVIDIA's CUDA framework.We translated existing functions to run in parallel on the GPU, starting with the functions where the most computing time is spent.Running on NVIDIA Tesla GPUs, initial results show an average speedup of 1.8 that stays fairly constant with increasing problem sizes.Though many applications can see even orders of magnitude improvement from GPGPU acceleration, the potential improvement for CitcomS is limited by several factors, including being bound by MPI collective communication.With newer GPGPU frameworks such as Fermi, further performance improvements can be expected, though a more significant overhaul of the CitcomS code would be necessary for any significantly better speedup.