SIParCS 2018- Akira "Aiko" Kyle
WRF Scaling, Performance Assessment and Optimization
The Weather Research and Forecast (WRF) model is a parallel mesoscale numerical weather forecasting application used in both operational and research environments. The performance of WRF on NCAR's Cheyenne supercomputer was investigated, focusing primarily on run time and compile time settings. The latest Intel (18.0.1) and Gnu (8.1.0) compilers were compared with different compilation flags. We found the Intel compiler to be consistently faster than Gnu at various optimization levels. Various MPI libraries were tested, including MPT, MVAPICH, Intel MPI, MPICH. We found that openMPI, MPT, and MVAPICH showsimilar runtime performance while the performance of MPICH was poor and Intel MPI's performance scaled poorly to large node counts. Several benchmark cases were developed for the latest version of WRF (4.0) at different resolutions and utilizing the CONUS and tropical physics suites. The scaling results of these benchmark cases were used to give users of Cheyenne recommendations on how to run WRF in a timely and efficient manner. Additionally, we compared the scaling of WRF's initialization time, I/O time, and computation time and found better scaling to very large node counts than with previous WRF versions on NCAR's previous supercomputer, Yellowstone.
Mentors: Davide Del Vento, Brian Vanderwende, Negin Sobhani