文摘
This research focuses on the development of techniques to improve the performance of Computational Fluid Dynamics (CFD) calculations in a parallel computing environment. In particular, this research develops algorithms to handle CFD computations on unstructured grids. Many scientific computations involve dividing the computational domain into a large number of small elements (cells) to form a grid which may be structured or unstructured. The structured grid is simple to implement but too inflexible in the cell shape for complex flow fields. Unstructured grids are an attractive alternative because they allow great flexibility in the types of cells, but there have been obstacles to their efficient implementation. Parallel computing can be a cost-effective way of improving the efficiency of large scale CFD computations, but the nature of the grids and solution methods present obstacles to getting the full benefits from parallelization. This research develops new methods to resolve several challenges in CFD calculations. Specifically, we have successfully developed and tested new parallel techniques for (a) dynamically refining and coarsening unstructured grids with different shapes of cells (hybrid grids), (b) distributing the work load evenly amongst processors (dynamic load balancing), (c) creating a parallel version of the Algebraic Multi-grid (AMG) equation solver, and (d) predicting the flow fields around multiple moving objects using multiple, overlapping grids that move with the objects. Extensive tests on 2D and 3D complex flow problems have verified that our new methods have accomplished (1) efficient refinement and coarsening of hybrid grids with improved solution accuracy and speedup, (2) improved work load distribution amongst processors, (3) significant acceleration of the equation solver, and (4) accurate and fast prediction of the flow fields with multiple moving objects. Future research will develop effective algorithms to further control the communication cost and distribute work load evenly.