Parallel grid preprocessing method based on mpi

Abstract

The invention discloses an MPI-based grid parallel preprocessing method, including the grid partition number of a given computing domain; starting MPI multi-processes, setting the number of processes; judging whether the number of processes is equal to the number of partitions, and opening the network Grid file, the main process reads the grid unit information file, and initially distributes the grid units to each process on average, and each process creates its adjacency array, otherwise restarts the MPI multi-process; each process calls ParMETIS for grid division; Each process reads the grid unit information into the array in blocks, and sets the index position of the array; each process loops through the grid unit information file, and judges whether the length of the array minus the index position number of the array is less than the length of the grid unit information, less than Then read the data of the grid unit information file and fill it into the array, otherwise, assign the array element to the grid unit; judge whether the partition number of the grid unit is equal to the process number, if it is equal, store the grid unit information in the process file, Otherwise, continue to loop to judge.

Description

technical field [0001] The invention relates to a parallel preprocessing technology, in particular to a grid parallel preprocessing method. Background technique [0002] In the field of scientific and engineering computing, the grid is of great significance to the numerical solution of various differential equations, and the grid distribution is the basic environment for solving calculations. The solution of differential equations mainly includes two steps of numerical discretization and solution of algebraic equations. When the discretization method is determined, the grid distribution information can directly reflect the logical structure of algebraic equations solution vector and coefficient matrix. With the widespread application of parallel computing, grid plays a very important role in the parallel solution of differential equations. For distributed parallel computing, the parallel execution of grid division based on domain decomposition and distributed storage of gri...

AI Technical Summary

Problems solved by technology

[0004] Data segmentation and grid information management are the main time-consuming stages of grid parallel preprocessing. The existing technology is time-consuming and inefficient in the data segmentation and grid information management stages of grid preprocessing.

For grid division, the existing multi-layer recursive bisection, row and column division and other technologies are slow and the division quality is not ideal

Most of the existing grid preprocessing schemes are serially executed, which can only be executed on a single CPU core. At the same time, most of the grid files are traversed serially, and the speed is relatively slow.

Moreover, existing grid preprocessing schemes mostly store grid data files in one or a few files. When the data scale is large, I / O file read and write congestion will occur, which affects the scale of grid processing data. and speed

Images