Design And Implemention Of PETSc's Heterogeneous Parallel Algorithm On SW26010 Processor

With the continuous development of science and technology,all countries in the world are making great efforts to develop their own supercomputers,and the performance ranking of supercomputing also shows the comprehensive national strength of a country.In recent years,the performance ranking of supercomputers in China has always ranked first in the world,and can not be separated from the strong support and input from the state.Under this premise,the national parallel computer engineering technology research center independently developed the Shen Wei multiprocessor and built the world's best performance supercomputer,the shinwei Taihu light.The first work of this article is to research and study the architecture of the Deuteronomy nuclear processor,and to the total of the the Sunway TaihuLight.The structure of the body is understood.The PETSc toolbox,supported by the US Department of energy,is widely used in solving large-scale partial differential equations.It is the base for large-scale scientific numerical computation.In many applications of supercomputing,common common mathematical libraries need to be called,in which the PETSc function library is the kind of Mathematical Library.The PETSc function library has been fully used in the CPU+GPU heterogeneous system,but the application base of the domestic heterogeneous core isomeric chips is zero.It is necessary to realize its homemade heterogeneous core isomeric chip.First,we need to understand the software structure of PETSc function lib rary deeply.So the second task of this thesis is to learn and research PETSc function library.After the study and research of hardware and software,the third work of this thesis is to implement the entire PETSc function library based on the features of software and hardware,and optimize the core functions of the part of the PETSc function library according to the needs of the research.At the time of operation,the most use of the 8192 computing nodes of Sunway TaihuLight is used.The experimental results show that the parallel algorithm of some core functions is optimized,the acceleration ratio of the single node is up to 16.4,and the theoretical acceleration ratio is basically reached.In the multi node MPI scalability test with large input scale,the running time and the 256 of the 8192 nodes are run and 256.The ratio of the running time of the node is 32,and the acceleration ratio increases linearly with the linear increase of the number of heterogeneous processors,indicating that the PETSc core function parallel algorithm has good scalability on the optical supercomputer of the Sunway TaihuLight.