Study on Massively Parallel Computational Mechanics

为了克服向量流水线、共享内存型超级计算机对计算能力的限制，超级计算机的发展趋势是向分布式存储的大规模并行平台发展。本文的研究目标是能够充分利用大规模并行处理器的高性能计算。研究课题包括两个方面：一是将传统的和/或现有的软件应用到大规模并行环境中，获得高可伸缩性；二是开发适合大规模并行平台的创新算法。首先，在CFD(计算流体动力学)领域，将基于EBE(Element-by-Element)的CFD程序移植到KSR1或T3D上，观察到EBE程序内在的并行性。提出了一种非结构化多色算法，并在带反馈机制的神经网络求解器中实现了该算法。神经解算器是US…其次，提出了一种无网格方法，该方法只需要节点数据进行计算，避免了繁琐的网格生成过程。与传统的基于Delaunay三角剖分的有限元方法进行了比较，分析了该方法的精度和效率。文中还考察了无网格法对大规模并行平台的适用性。对无网格法进行了进一步改进，实现了自适应分析和流体分析，充分利用了无网格法的特点。作为无网格法的另一种形式，提出了一种并行合作型有限元方法。结合人工生命即智能体的思想，实现了协作式网格自动生成。最后，提出了连续空间搜索中的进化算法，并将其应用于本构方程参数识别这一反问题中。并对进化算法的并行性进行了研究。较少

For overcoming the limitation of the computational power by the vector-pipeline, shared memory-type supercomuputer, trend in the supercomputer development is towards the massively parallel platform with the distributed memory. This study aims at the high-performance computing, which can make the best use of the massively parallel processors. Research subjects consist of two approaches ; one is to apply the conventional and/or currently available software to the massively parallel environments and to attain the high scalability, and the other is to develop the innovative algorithms, which are suitable for the massively parallel platform.First, in the CFD (Computational Fluid Dynamics) area, a CFD code, which is constructed on an EBE (Element-by-Element) basis, is ported to KSR1 or T3D.The inherent parallelism of the EBE code was observed. Also, an unstructured multi-color method is proposed, and implemented in the neural network solver with the feedback mechanism. The neuro-solver is us … More ed to solve the Poisson's equation for pressure.Secondly, a meshless method is proposed, which needs only nodal data for computations, and hence avoids burdensome mesh generation process. Comparing with the conventional FEM,which utilizes the Delaunay triangulation, its accuracy and efficiency are investigated. The applicability of the meshless method to the massively parallel platform is also examined. The meshless method is furthermore enhanced to realize the adaptive analysis and the fluid analysis, both of which can take good advantage of the meshless nature. As another type of meshless method, a parallel cooperationtype type FEM is proposed. Incorporating the idea of the artificial life, i.e.agent, cooperationtype type automatic mesh generation is realized.Finally, the evolutional algorithm in the continuum space search is proposed, and is applied to an inverse problem, i.e.the parameter identification for constitutive equations. Parallel nature of the evolutional algorithm is also investigated. Less