Methods and systems for optimizing sparse matrix vector multiplication of a high-performance computing framework

Abstract

The invention discloses a method for optimizing sparse matrix vector multiplication of a high-performance computing framework. According to the method, sparse matrix vector multiplication operation isdivided into column multiplication operation and row addition operation: the column multiplication operation is firstly carried out according to columns of a sparse matrix, and in the process, data access to an input vector x is changed from irregular and discrete to continuous; in the process, the data access of the output vector y is changed from irregular and discrete to continuous, so that the problem of high memory access delay caused by irregular data access is avoided; the invention further designs a four-layer division mechanism, which comprises kernel group layer division, customizeddivision, slave kernel layer division and local memory layer division, so that the multi-level calculation framework and the memory structure of the Sunway Taihu Light are fully utilized, and the problems of local memory limitation and load imbalance on the calculation kernel are avoided.

Description

technical field [0001] The invention belongs to the field of parallel computing, and more specifically relates to a method and system for optimizing sparse matrix-vector multiplication of a high-performance computing framework. Background technique [0002] At present, the high-performance computing architecture has been increasingly common in industrial applications. Among them, the "Sunway Taihu Light" based on the SW26010 multi-core heterogeneous processor is a typical high-performance computing architecture. It is developed by the National Parallel Computer Engineering Technology Research Center Independent research and development, the supercomputer installed in the National Supercomputing Wuxi Center is equipped with 40960 SW26010 processors. There are 4 core groups on each SW26010 processor, and each core group is installed with a master core and 8*8 slave cores, in which the master core is responsible for preprocessing, slave core computing task allocation, and some ...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the above defects or improvement needs of the prior art, the present invention provides a method and system for optimizing the sparse matrix-vector multiplication of the high-performance computing framework. The technical problem of high memory access delay in the parallel sparse matrix multiplication operation, and the technical problem that it cannot handle large-scale sparse matrix operations because of the memo memory on the slave core is only 64KB, which affects its large-scale application. The sparsity of its matrix leads to technical problems of unbalanced load on slave cores

Images