Análisis de rendimiento y profiling del modelo WRF en un clúster HPC

Abstract

The Weather Research and Forecast (WRF) model is a fully functional modeling system that supports atmospheric research and weather prediction. WRF was developed with emphasis on efficiency, portability, maintainability, scalability and productivity, allowing it to be successfully implemented in a wide variety of HPC equipment. Therefore, the size of the problems supported by WRF has increased, so the understanding of the WRF’s dependence on the various elements of the cluster, such as CPU, networking and libraries are crucial to enable efficient forecasting and high productivity. In this context, this manuscript examines WRF scalability in HPC equipment, taking into account three parameters: number of CPUs and nodes, communications and libraries. Two benchmarks carried out on a cluster of high performance provided with a GigaEthernet network, allow to establish the relationship between scalability and the three parameters studied, particularly WRF demonstrates sensitivity to inter-node communication. This factor is essential for maintaining the scalability and increasing productivity by adding nodes to the cluster.