Parallel Unstructured Mesh Generation Approaches For Large-Scale Numerical Simulations

In the field of high performance computing (HPC), it is an unarguable reality that numerical simulation systems lags behind HPC machines far away in terms of the parallel capability. Here, a key factor that slows down the parallel capability of the state-of-the-arts numerical simulation systems is the sequential implementation of the technologies that enable a numerical simulation, such as mesh generation. In fact, many important large-scale simulations still depend on a sequential meshing procedure nowadays, although the solvers may be able to run in a massively parallel computer efficiently. Not surprisingly, it remains a main performance bottleneck to create a mesh for simulations that involves hundreds of millions of freedoms. Thus, it is a key approach to overcome such a bottleneck by developing parallel meshers that can run on massively parallel computers.This study focuses on problem parallel approaches for mesh generation approaches. It achieves the following contributions in the aspects of both sequential meshing and domain decomposition.Firstly, a new operation is proposed to remove a point from a tetrahedral mesh by the small polyhedron reconnection technique. This operation successes in more cases than those existing point removal operations mainly because of its exhaustive search nature. The number of Steiner points required for boundary constraint tetrahedral mesh generation can then be reduced remarkably by suppressing Steiner points using this operation. As a result, the reliability of tetrahedral mesh generation is enhanced, and the quality of output elements is more likely to be ensured.Secondly, a parallel tetrahedral mesh generation approach based on recursive bi-divisions using triangular surfaces is proposed. Research is conducted for addressing issues concerning the reliability of domain decomposition. A novel procedure employing local modification techniques is proposed for repairing the intersecting inter-domain mesh instead of directly repeating the bi-division procedure, which improves the robustness of the domain decomposition procedure and thus the complete meshing procedure significantly.Thirdly, a mesh simplification approach is proposed by analyzing the dual graphs of the input mesh, featured by its ability of removing undesirable mesh sides (edges in 2D or faces in 3D). The simplified mesh instead of the original mesh is partitioned using generic graph partitioner, which yields many subdomains without small dihedral angles and/or poorly shaped sides attached on the inter-domain boundary. This domain decomposition approach can facilitate the downstream applications of parallel mesh quality improvement and parallel mesh generation to be implemented in a fully code-reuse of sequential programs and highly efficient manner. The applications demonstrated in this study include a parallel surface mesher, a parallel volume mesher and a parallel mesh improver.The above contributions enable the development of two parallel preprocessing pipelines aimed at generating large-scale tetrahedral meshes in a completely parallel fashion. In other words, all of the modules involved in both pipelines such as surface meshing, volume meshing, volume mesh improving and domain decomposition are parallelised in both pipelines. Experimental results show that one billion tetrahedral elements that fill in the outflow domain of typical aircraft model can now be generated in less than 7 minutes by using 256 computer cores. Presently, a completely parallel simulation environment has been developed by connecting the developed preprocessing pipelines and an in-house computational aerodynamics code. Meshing and simulating experiments are performed to demonstrate the reliability, effectiveness and efficiency of the developed simulation environment and its potential applications to large-scale simulations of complex aerodynamics models.