Development and performance optimization of a parallel computing infrastructure for an unstructured-mesh modelling framework

Abstract. This paper describes the development and performance optimization of a parallel computing infrastructure for an unstructured-mesh global model (GRIST; Global-to-Regional Integrated forecast SysTem). The focus is on three major aspects that facilitate rapid iterative development, including parallel computing, index optimization and an efficient group I/O strategy. For parallel computing, the METIS tool is used for the partition of the global mesh, which is flexible and convenient for both the quasi-uniform and variable-resolution simulations. The scaling tests show that the partition method is efficient. To improve the cache efficiency, several mesh index reordering strategies are investigated to optimize the performance of the indirect addressing scheme used in the stencil calculations. The numerical results show that the indexing strategies are able to speed up the calculations, especially for running with a small number of processes. To overcome the bottleneck of poor I/O efficiency for the high-resolution or massively parallel simulations, a group parallel I/O method is implemented and proven to be of high efficiency in the numerical experiments. Altogether, these three aspects of the parallel computing toolkits are encapsulated in a few interfaces, which can be used for general parallel modelling on unstructured meshes.