Scalability Analysis of Weather Research Forecast Model on NVIDIA Ampere based Dense GPU Cluster

WRF (Weather Research Forecast) is one of the commonly used weather applications for operational and research purposes. WRF’s GPU accelerated version is being used to have higher-resolution model forecasts for better accuracy, in accelerated time-to-solution. WRF’s GPU acceleration also provides deeper insights into upcoming weather phenomena with more time in operating decisions. This paper focuses on the performance and scalability analysis of the GPU accelerated WRF on the PARAM Siddhi-AI system. PARAM Siddhi-AI system has AMD EPYC Rome CPUs and NVIDIA’s ampere architecture-based (A100) GPUs. The NVIDIA Ampere architecture, designed for the age of elastic computing, rises to all these challenges, providing unmatched acceleration at every scale. The faster and larger L1 cache and shared memory in Ampere architecture (192 KB per Streaming Multiprocessor (SM)) delivers additional speedups for High-Performance Computing (HPC) workloads. Accelerated WRF performance was analyzed for two input datasets namely easter500 (small input with 12.7 million grid points) and for the user-defined (large input with 93.5 million grid points) domain, covering four continents e.g., Asia, Europe, Africa, and Australia respectively. GPU-enabled WRF’s scalability is analyzed on single-node as well as on multiple nodes (up to 5 nodes). The performance of the accelerated WRF is compared with its CPU-only version as well. For easter500 input, the relative speedup observed by accelerated WRF is 6.7x in comparison to WRF CPU-only performance. In the case of a user-defined (larger) input data set, the performance gain from accelerated WRF single-node run and multiple nodes run is 25x and 46x respectively.