Approaching Exascale with the Weather Research and Forecasting Solar Model

We investigate the scalability and I/O performance of the Weather Research Forecasting (WRF) model on the PRACE Tier-0 systems SuperMUC-Haswell, Marconi KNL, and Marenostrum 4 within the PRACE 5th Implementation phase in cooperation with the Energy Oriented Centre of Excellence (EoCoE). We identify the bottlenecks which can impede WRF to run eciently on pre-exascale and exascale machines. We report on the performance of both the computational kernels and I/O as captured using Scalasca/Score-P and extrae. From a series of strong scaling tests and benchmarks in which we vary the parallelization using di erent MPI and OpenMP setups, we identify strategies to run eciently on PRACE Tier-0 machines. The analysis is carried out for three schemes, namely the weather forecast scheme, the scheme based on WRFsolar, and the WRF- re scheme