Performance Analysis, Profiling and Optimization of Weather Research Forecasting (WRF) Model

The Weather Research and Forecasting (WRF) model is an open source mesoscale numerical weather prediction (NWP) system extensively used in atmospheric research, operational forecasting and educational settings. A large high-resolution case of Hurricane Sandy has been used for benchmarking and scaling analysis of WRFV3.7 on Texas Advanced Computing Center (TACC) Stampede cluster. MPI and hybrid (MPI-OpenMP) parallelization of WRF were tested changing the numbers of MPI tasks and OpenMP threads.

We used TAU (Tuning and Analysis Utilities) and Intel Vtune Amplifier XE for profiling and identifying hotspots and bottlenecks of WRF. The results show that advection is the most time consuming routine in the ARW (Advanced Research WRF) dynamic solver. For this benchmarking WRF 5th order positive-definite advection scheme were used. Positive definiteness, which prevents negative concentrations, is used for conserving mass during the model runtime. Further manually instrumented code shows that 5th order positive definitive advection can be optimized for higher efficiency.