SFC-based Communication Metadata Encoding for Adaptive Mesh Refinement

Abstract

The present paper studies two adaptive mesh refinement (AMR) codes whose grids rely on recursive subdivison in combination with space-filling curves (SFCs). A non-overlapping domain decomposition based upon these SFCs yields several well-known advantageous properties with respect to communication demands, balancing, and partition connectivity. However, the administration of the meta data, i.e. to track which partitions exchange data in which cardinality, is non-trivial due to the SFC's fractal meandering and the dynamic adaptivity. We introduce an analysed tree grammar for the meta data that restricts it without loss of information hierarchically along the subdivision tree and applies run length encoding. Hence, its meta data memory footprint is very small, and it can be computed and maintained on-the-fly even for permanently changing grids. It facilitates a fork-join pattern for shared data parallelism. And it facilitates replicated data parallelism tackling latency and bandwidth constraints respectively due to communication in the background and reduces memory requirements by avoiding adjacency information stored per element. We demonstrate this at hands of shared and distributed parallelized domain decompositions.