Research On Program Parallelization Method Based On Deep Learning

With the rapid development of high-performance computing technology,parallel programming technology plays an increasingly important role in practical engineering applications.Especially in the field of scientific computing,parallelization technology has become a key way to solve the problems of complex program structure,large amount of calculation,data intensive and long execution cycle.However,the existing program parallelism analysis tools still have greater limitations when processing programs with the above characteristics,and the current multiple parallel programming methods lack a unified standard programming interface when used.Therefore,the parallelization of sequential programs has great difficulty in both the parallelism discovery and parallel program development.It is necessary to do a research on parallelization method of sequential programs which has good performance in the above two stages.The purpose of this paper is to explore the parallelization method of sequential programs with the help of deep learning technology,focusing on two aspects of sequential program parallelism discovery and parallel program development,and proposes a set of solutions including data sets,recognition methods,markup languages and auxiliary platforms.The main work is as follows:(1)A sequential program parallelism discovery method based on deep learning is proposed.In this paper,the sequential program parallelism discovery is regarded as a binary classification problem.A deep learning model based on deep graph convolutional neural network(DGCNN)is established,and a general graph data set GFCPD(Graphs For Code Parallelism Discovery)is constructed.The performance of the model on the GFCPD data set verifies that the deep learning method has feasibility and effectiveness in the sequential program parallelism discovery problems.From a new perspective,a new solution is proposed for the sequential program parallelism discovery programs.(2)Design and implement a parallel markup language PML and programming assistant platform.This paper uses XML technology to implement the parallel markup language PML which provides unified formatted PML tags for different parallel programming models.Providing a unified standard programming interface and a simpler and more convenient way for the mixed use of a variety of parallel programming methods.In addition,this paper builds a parallel programming auxiliary platform,which provides a complete sequential program parallelization process from the three aspects of serial program parallelism discovery,parallel program development,and parallel program debugging.It has lowered the threshold of parallelization technology.