Performance Evaluation of the Weather Research and Forecasting (WRF) Model on the DOE Summit Supercomputer

A series of experiments are performed on Summit supercomputer to benchmark and characterize the computational performance of the Weather and Research Forecasting (WRF) model on CPU and GPU platforms. The implementation includes the complete dynamical core and select physics options based on OpenACC directives, such that a full WRF model can be executed on GPU-based systems. Previous studies have shown that new computing platforms, such as GPUs, have substantial speed-ups over the traditional CPU based computations. Our benchmarks indicate that the model can perform up-to 8x in dynamics-only and 5x faster in physics-enabled simulation configurations with GPUs. For average loads, 1 GPU is about 2 times faster than the entire CPU node on Summit supercomputer. Memory usage statistics of CPU and GPU setups were also explored in single and multi-node parallel configurations. Memory footprint of a simulation domain grows linearly with increasing rank counts which points inefficiencies in the model’s memory management. Additionally, we will present a comparison of variation in track and intensity of Hurricane Sandy to assess reproducibility between the platforms. We plan to extend benchmark efforts to include fine-scale model timings, such as dynamics solver components, MPI specifics, and IO layer. Discussions in-progress to port domain nesting capability and urban canopy specific model functionality in the GPU model. Our benchmark results highlight that high-resolution NWP simulations require substantial algorithmic redesign and/or surrogate models to accelerate time-to-solution in order to efficiently utilize exascale computing resources.