Dynamic Adaptive Refinement In Earth System Modelling

Abstract

: Increasing the mesh resolution is one the most important tools for increasing the accuracy of numerical simulations. However, increasing the mesh resolution globally increases the amount of data and computational time substantially. In the exascale-era sub-km meshes are becoming more and more popular for atmospheric models, making it especially difficult to manage the vast amount of data efficiently. With dynamic adaptive mesh refinement (AMR) we locally control the resolution of a mesh in areas of interest, using a fine resolution only where it is explicitely needed and keeping the mesh coarse elsewhere. Thus, we concentrate the data and computing power and significantly reduce the simulation costs while keeping the same numerical accuracy. Vice versa, the resolution can be increased while keeping the same runtime. Managing adaptive meshes induces new challenges such as load-balancing, mesh management, ghost layer computation etc. Developments in the recent years have extended the scalable and efficicient tree-based AMR approach from quadrilaterals/hexahedra to various element shapes such as triangles, tetrahedra, pyramids or prisms and have been implemented in our AMR library t8code. It is a third-party library that adresses these challenges and can be integrated by solver environments in order to enable AMR. In our presentation we will give an introduction to AMR and how we use it for atmospheric simulations. We will give an overview of ongoing and past projects, such as our contributions to PilotLab Exascale Earth Sytem Modelling (Pl-ExaESM) or the lossy data compression for data coming from atmospheric simulations. Furthermore, we will demonstrate the efficiency of our methods with recent benchmark results on current supercomputers, showing that t8code scales on up to at least 1 million MPI ranks and over 1 Trillion mesh elements.