E-Thesis 87 views 35 downloads
Unstructured parallel grid generation. / Rajab Said
Swansea University Author: Rajab Said
PDF | E-ThesisDownload (16.5MB)
The ultimate goal of this study is to develop a 'tool' by which large-scale unstructured grids for realistic engineering problems can be generated efficiently on any parallel computer platform. The adopted strategy is based upon a geometrical partitioning concept, where the computational d...
No Tags, Be the first to tag this record!
The ultimate goal of this study is to develop a 'tool' by which large-scale unstructured grids for realistic engineering problems can be generated efficiently on any parallel computer platform. The adopted strategy is based upon a geometrical partitioning concept, where the computational domain is sub-divided into a number of sub-domains which are then gridded independently in parallel. This study focuses on three-dimensional applications only, and it implements a Delaunay triangulation based generator to generate the sub-domain grids. Two different approaches have been investigated, where the variations between them are limited to (i) the domain decomposition and (ii) the inter-domain boundary gridding algorithms only. In order to carry out the domain decomposition task, the first approach requires an initial tetrahedral grid to be constructed, whilst the second approach operates directly on the boundary triangular grid. Hence, this thesis will refer to the first approach as 'indirect decomposition method' and to the second as 'direct decomposition method'. Work presented in this thesis also concerns the development of a framework in which all different sub-algorithms are integrated in combination with a specially designed parallel processing technique, termed as Dynamic Parallel Processing (DPP). The framework adopts the Message Passing Library (MPL) programming model and implements a Single Program Multiple Data (SPMD) structure with a Manager/Workers mechanism. The DPP provides great flexibility and efficiency in exploiting the available computing resources. The framework has proved to be a very effective tool for generating large-scale grids. Grids of realistic engineering problems and to the order of 115 million elements, generated using one processor on an SGI Challenge machine with 512 MBytes of shared memory, will be presented.
College of Engineering