Heterogeneous parallel computing method for row updating of sparse matrix LU factorization

Abstract

The invention relates to a heterogeneous parallel computing method for updating sparse matrix LU decomposition rows. Based on the superlu algorithm, combined with the master-slave heterogeneity of Taihu Lake Light, the superlu solver transfers the large-scale computing task to the slave kernel in the process of matrix decomposition and row updating, and utilizes the efficient computing and data communication ability of the slave kernel to improve the computational ability of the super-large-scale sparse matrix and the overall performance of the solution. The invention divides tasks according to the data of the matrix, Firstly, a matrix space is created for each slave kernel, and each slave kernel is responsible for solving several matrix rows. In the process of solving the linear equationsof the lower triangular identity matrix type, each row of data is independent of each other, avoiding the conflict and dependence of the solution space, and smoothly realizing the solution of the equations.

Description

technical field [0001] The invention relates to a heterogeneous parallel computing method for row update of sparse matrix LU decomposition, and belongs to the technical field of sparse matrix LU decomposition. Background technique [0002] In recent years, my country has made a series of significant progress in fields closely related to electromagnetism. Among them, the numerical calculation of electromagnetic field is playing an increasingly important role because of its significant advantages such as high efficiency, flexibility, and convenience. In the numerical analysis method of electromagnetic field, moment Quantitative method has the characteristics of high theoretical precision. Moment method converts the electromagnetic field operator equation to be solved into a matrix equation. Due to its high theoretical precision, when dealing with electromagnetic problems of complex electric and large systems such as airborne phased array The huge complex number sparse matrix pre...

AI Technical Summary

Problems solved by technology

In the row update stage, superlu will perform a large number of unit lower triangular matrix equations solving operations, and the efficiency of this type of solving process is not high when calling the blas library provided by the usual TaihuLight, which is much lower than the simple main core The operating efficiency of the , resulting in a computational bottleneck

Images