Porting and Tuning WRF Physics Packages for Intel Xeon Phi and NVIDIA GPU

Two emerging non-hydrostatic global NWP models, NCAR's MPAS and NOAA's NIM, share a subset of physics packages from the well-known Weather Research and Forecast model (WRF). Since NIM and MPAS both target sub-10km scales, there is much interest in making efficient use of fine-grained computer architectures to reduce model execution times. Work is now underway to tune performance of the shared physics packages on Intel Xeon Phi and NVIDIA GPU. This talk will cover details of single-node performance tuning on Intel Xeon Phi and efforts to retain single source code across CPU, GPU, and Phi. GPU results will also be presented. Design tradeoffs involving performance, threading, vectorization, and long-term maintainability will be examined.