A sample of my work for the high performance computing group at the Space-Science and Engineering Center (SSEC), at UW-Madison.
The Weather Research and Forcasting model is a numerical weather prediction model used for forcasting and running climate models for research.
It's Runge-Kutta numerical integration on huge parallel arrays, which is super expensive computationally, but ideal for GPGPU. (i.e. embarrassingly parallel)
The original code is in Fortran-90, but my group was trying to use CUDA-C to do the hard parts on Nvidia GPU-clusters. I worked on translating a significant portion of the code to C, and rewrote most of the computationally difficult portions to run as massively-parallel algorithms. It's a game of micro-seconds, because it all adds up.
This is a sample of the work. Fortran, C, and CUDA versions of the advance_mu_t module, which is a small dynamics module.
On a desktop with 3 GTX-680 GPUs:
| Version of module | Time |
|---|---|
| Original Fortran | 152.0 ms |
| CUDA-C version | 0.051 ms |
#####3235x speedup! ooh-e!
The idea is to use small desktop GPU-clusters to run simulations in a few minutes, that would have otherwise taken hours without a mad-expensive CPU-cluster.