Seeking the best Weather Research and Forecasting model performance: an empirical score approach

Weather forecasting, especially snowfall prediction, was critical in the 2018 Winter Olympics, where the accuracy of the predictions was of key importance for the planning of the different Olympic events. It was a significant challenge for the authors to meet the requirements in time and forecast resolution, while doing their best to be as competitive as possible. All the forecasts were obtained using the Weather Research and Forecasting (WRF) model, executed on the GALGO supercomputer. In order to obtain the best performance and meet the required execution times, different combinations of compilers, Message Passing Interface (MPI) libraries and computing platforms were tested to seek the best combinations. This work proposes an empirical score of special interest to supercomputer maintainers, developers and scientists, which can be useful to obtain the best WRF configuration for their systems. Additionally, we found substantial performance differences when using different combinations of compilers, MPI libraries and hybrid shared memory paradigms, although these differences varied depending on the underlying platform. As conclusion, after all the tests we performed, we chose the combination with Intel compilers, Intel MPI library and OpenMP for the production system tasked to perform the weather forecasts for the Winter Olympic Games.