A parallel preprocessor applied to fluid dynamics problems

Abstract

As the computing power and memory resources of sequential and parallel computers continuously increase, the typical size of the meshes used for Finite Element methods in numerical analysis grows extremely large. The need for fast, good-quality mesh generators is crucial, since this preprocessing phase tends to become the performance bottleneck of a numerical application.

This paper presents a parallel preprocessor based Upon a partitioning approach for the generation of large-scale 3D unstructured meshes. The prepartitioner [3] takes as an input the physical domain boundary, described by a list of bounding facets Γ_i (issued by a CAD system for instance), that we will call a skin. It generates a separator of the domain into sub-domains, which is a list of facets that will be noted Γ_o Hence, each sub-domain is described by a skin included in Γ_i ∪ Γ_o. Then, each computational step, including the 3D mesh generation, is performed in parallel on each sub-domain.

Contrary to related work in building such separators, we consider that we have absolutely no knowledge of the inside of a domain, neither CAD information [1, 6] nor any underlying coarse mesh [8]. Only the description of the domain boundary (the skin) is known.

We applied our approach to N3S, the fluid dynamics package developed by EDF and distributed by Simulog [2], which has been parallelized within the Europort-1 consortium HPCN3S [5].