面向大规模数值模拟的并行非结构网格生成方法研究

发布时间：2018-06-16 13:47

本文选题：网格生成 + 数值模拟　；参考：《浙江大学》2016年博士论文

【摘要】：总体上,当前的高性能数值模拟软件的并行能力远落后于硬件可提供的最大并行能力。制约数值模拟软件并行能力提高的关键不仅在于求解器并行能力不足,更多情形下是由于数值模拟支撑软件并行能力的滞后。现有的并行数值模拟软件大都采用“串行网格生成+并行求解”的“准并行”计算模式,在生成很多复杂问题求解所需的大规模网格时存在严重的内存和时间性能瓶颈。开展并行网格生成方法研究,实现“并行网格生成+并行求解”的“全过程并行”计算模式,是提高数值模拟并行计算能力的关键所在。本文聚焦问题并行类并行非结构网格生成方法,系统研究了与之相适应的串行网格生成和区域分解方法,取得创新如下：(1)针对四面体网格局部编辑问题,提出了基于小空腔重连技术的点删除操作,利用小空腔重连技术的穷举机制提高点删除的成功率,继而将点删除操作应用于Delaunay网格生成的约束边界恢复算法中,显著减少了复杂输入情形下边界恢复算法所需Steiner点的数目,提高了四面体网格生成的可靠性和单元质量。(2)改进了一类基于表面网格递归分解的区域分解方法,提出一系列网格局部编辑操作以及它们的组合操作,有效解决了区域分解中间过程产生的子区域表面网格相交问题,提高了该类区域分解方法的计算效率和可靠性,继而实现了一类基于该区域分解方法的并行四面体网格生成算法。(3)提出一类基于网格对偶图的网格简化算法,可消除初始背景网格包含的不理想形状特征。继而提出一类以简化后网格为输入、基于图分解算法的新区域分解方法,基于它分解得到的子区域结构除满足负载平衡和通信最小化等目标外,且子区域边界的形状质量较高,减少了子区域边界对并行网格生成或优化算法所形成的单元质量的影响。基于该区域分解方法,实现了完全解耦的并行曲面网格生成、并行四面体网格生成及并行四面体网格优化算法。最终,针对不同问题需求,集成上述创新成果,开发了两类全过程并行非结构网格生成软件。以诸如飞行器整机外形这类复杂模型的CAD文件为输入,利用256个核在不到7分钟时间内实现了包含十亿实体单元的大规模计算网格的全自动、全过程并行生成。目前,上述并行软件已和自研的并行空气动力学数值模拟软件实现了无缝集成,针对典型空气动力学计算问题实现了“并行网格生成+并行求解”的“全过程并行”计算模式,解决了这些问题精细分析所面临的大规模计算网格生成难题。
[Abstract]:In general, the parallel capability of the current high performance numerical simulation software lags far behind the maximum parallel capability that hardware can provide. The key to the improvement of parallel capability of numerical simulation software is not only the lack of parallel ability of solver, but also the lag of parallel ability of numerical simulation software in many cases. Most of the existing parallel numerical simulation software use "serial grid generation parallel solution" quasi-parallel "computing model, there are serious memory and time performance bottlenecks in generating large scale grid for solving many complex problems. It is the key to improve the parallel computing ability of numerical simulation to carry out the research on parallel grid generation method and to realize the "whole process parallel" computing model of "parallel grid generation and parallel solution". In this paper, the parallel class parallel unstructured mesh generation method for focusing problem is discussed. The corresponding serial mesh generation and domain decomposition methods are systematically studied. The innovations are as follows: 1) for the local editing problem of tetrahedron mesh, The point deletion operation based on the small cavity reconnection technique is proposed. The exhaustive mechanism of the small cavity reconnection technique is used to improve the success rate of point deletion, and then the point deletion operation is applied to the constrained boundary restoration algorithm of Delaunay mesh generation. The number of Steiner points required for the boundary restoration algorithm in complex input cases is significantly reduced, and the reliability and cell quality of tetrahedron mesh generation are improved. The domain decomposition method based on recursive decomposition of surface mesh is improved. A series of local grid editing operations and their combination operations are proposed, which can effectively solve the problem of the intersecting of subregion surface meshes generated by the intermediate process of domain decomposition, and improve the computational efficiency and reliability of this kind of domain decomposition methods. Then a class of parallel tetrahedron mesh generation algorithm based on the domain decomposition method is implemented. A mesh simplification algorithm based on the grid-to-bipartite graph is proposed which can eliminate the undesired shape feature of the initial background mesh. Then, a new domain decomposition method based on graph decomposition algorithm and simplified grid as input is proposed. The subregion structure based on it not only satisfies the objectives of load balancing and communication minimization, but also has high shape quality of sub-region boundary. The effect of sub-region boundary on the cell quality of parallel mesh generation or optimization algorithm is reduced. Based on the domain decomposition method, the fully decoupled parallel surface mesh generation, the parallel tetrahedron mesh generation and the parallel tetrahedron mesh optimization algorithm are implemented. Finally, two kinds of parallel unstructured grid generation software are developed to meet the needs of different problems and integrate the above innovative results. With the input of CAD files of complex models such as aircraft shape, 256 cores are used to realize the full automatic and parallel generation of large scale computing meshes containing 1 billion solid units in less than 7 minutes. At present, the above parallel software has been integrated seamlessly with the self-developed parallel aerodynamics numerical simulation software, and the "whole process parallel" computing mode of "parallel mesh generation and parallel solution" has been realized for typical aerodynamics computing problems. The problem of large scale mesh generation in fine analysis is solved.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP338.6

【相似文献】