Memory-Optimized Tile Based Data Structure for Adaptive Mesh Refinement

Abstract

Multi-scale simulation is relevant for many applications, such as modelling of fluids, electromagnetic or seismic waves, plasma physics, and it stands on the borderline of the supercomputer abilities. For this kind of problems, the Adaptive Mesh Refinement (AMR) methods aim to provide higher cell resolution only in areas, where it is necessary, while these domains may change in time. We propose a new framework for AMR data structure, with the goal to minimize the memory overhead for data storage, and, at the same time, to optimize the locality of data access. With higher locality, the performance gain of the computation is achieved by the use of the faster memory for each parallel processor. In the proposed framework, the cell data is combined in tiles. Two type of tiles (light and heavy) are used for minimizing the memory overhead in case a tile is sparsely filled. The interface allows implementation of various numerical methods. It provides a choice for a traversal rule with an iterator structure, which may be used for algorithms with higher operational intensity. The dynamic mesh adaptation works well for meshes that cover complex geometry.