PETSc Analysis And Some Design Considerations Of Parallel Numerical Software For High Performance Computing

As applications grow in size, together with the development of computing technology, parallel computing gets more and more necessary and has become the only way to solve many large-scale scientific computing problems.The application of parallel computing cannot exist without the support of related software. However, the design of a high-performance parallel program usually requires much more efforts than a sequential one, especially in case of complex problems. Therefore, there is an urgent need of the development and improvement of high-performance numerical software.Developing high-performance numerical software is a difficult task. It needs deliberate efforts from design to implementation, which often takes a long period of time. Under this condition, obtaining technical and empirical knowledge through understanding, using and analyzing some existing well-designed software has become a shortcut to achieve our goal.The Portable, Extensible Toolkit for Scientific Computation, PETSc, has become a model of the high performance numerical software which gains huge attention and wide use in commuting community throughout the world. Our selection of PETSc as the target model of studying is based on following reasons:PETSc uses MPI for all parallel communication, which is most fittable for general scalable computing, especially on our PC-Cluster platform. PETSc is a general purpose suite: of tools for the scalable solution of partial differential equations and related problems, which is much in accord with our research focus. Particularly, the established form, efficiency and methodology and principles of PETSc, give us a proper direction in numerical software development which is extremely important.With the explicit intention, the main contributions of this dissertation are: Give a general description of the research background. The main research content are classified as numerical modeling, programming and software design.Probe into the internal organization of PETSc and do lots of software analysis. Some crucial design principles, methods and techniques in PETSc are highlighted.Propose a general scheme for the parallelization of an unstructured mesh, including graph partitioning, data division and mesh managing, based on message passing programming style.Propose a series of comprehensive considerations about how to achieve a well-designed, object-oriented and data-distributed parallel software.