Decomposition of a Surface Unstructured Computational Mesh for Scaling Computations on a Supercomputer

Abstract

Currently, when solving complex computational problems of computer modeling, computational grids containing tens and hundreds of millions of cells are used. To perform computations of such amount, it is required to use supercomputer clusters consisting of many computational nodes connected by a high-speed communication network. In this case, it is necessary to decompose the computational mesh into separate domains. These domains are distributed among the computational nodes of the supercomputer and are processed in parallel, independently of each other. To synchronize computations, after each iteration of cell processing, data exchanges are performed at the boundaries between adjacent contiguous domains. To efficiently perform computations and scale them to a large number of computational nodes, it is necessary to develop efficient algorithms for decomposition of computational meshes that generate many domains with imposed requirements for the number of cells, the uniformity of the distribution of cells across domains, the connectivity of domains and the size of boundaries between them. The article considers an unstructured surface mesh as an object of research, which is used to calculate the processes of interaction of a volumetric body with the environment. For a mesh of this type, various classes of decomposition algorithms are considered, and a hierarchical decomposition algorithm is proposed with the choice of the optimal criterion for partitioning domains. The proposed algorithm is implemented in the program code, the results of comparing the operation of this algorithm on test meshes with algorithms of other classes are presented.

Keywords

• supercomputer
• surface unstructured computational mesh
• decomposition
• domain
• high performance computing
• computation scaling