Author(s)

Yafang Li

Corresponding Author

Yafang Li

Abstract

This paper focuses on the research of heterogeneous high-performance system algorithm based on deep learning (DL) model. With the development of DL, heterogeneous high-performance system becomes the key to improve computing efficiency. The purpose of this study is to design a DL algorithm model adapted to heterogeneous environment. By analyzing the related theories and technologies of DL and heterogeneous systems, an algorithm model including task analysis, resource allocation, data flow optimization and execution coordination modules is constructed. The convolution neural network is tested on heterogeneous platform, and compared with the CPU-only running model, the training time of heterogeneous system using this algorithm model is significantly shortened, from an average of 45.6 minutes to 12.3 minutes, and the classification accuracy is improved from 78.5% to 85.2%. The results show that the proposed algorithm model can effectively utilize the advantages of heterogeneous systems, improve the classification performance while accelerating the training speed of the model, and provide an effective way for DL algorithm optimization in heterogeneous high-performance systems.

Keywords

Deep Learning Model; Heterogeneous High Performance System; Performance Optimization