Distributed Parallel Optimization Of Tensor Networks Based On Cloud Computin

Tensor network is a tensor decomposition method,which is widely used in data mining,image processing,multi-modal information fusion and so on.In recent years,with the development of “Cyber-Physical-Social Systems”(CPSS),the scale of tensor data,the main carrier of multi-modal data has grown rapidly.It is difficult to analyze and process large-scale multi-modal tensor data in the traditional single-machine processing process.Therefore,it is urgent to study the analysis and processing technology of tensor large data.Cloud computing platform is an effective way to solve the problem of tensor data processing because of its storage and computing capability of large-scale tensor data.This paper analyzes the development of tensor theory and tensor data analysis methods.By studying distributed file storage technology,Map Reduce parallel programming model,Spark parallel computing and other related cloud computing technologies,it designs a large-scale tensor data storage and reading method based on HDFS.Hierarchical tensor decomposition parallel optimization and tensor ring decomposition parallel optimization based on cloud computing are proposed.At the same time,this paper studies the management and application of tensor data in cloud computing environment and designs the management application of tensor data in cloud platform.This paper mainly includes:1、In-depth analysis of the parallel storage of tensor data,a distributed and parallel method for Hierarchical Tucker tensor decomposition based on cloud computing is proposed.This paper designs a tensor data storage method suitable for distributed environment,according to the characteristics of large-scale tensor data.Two levels of parallel computing between nodes and parallel computing within nodes are designed to solve the bottleneck of single machine,based on the characteristics of hierarchical tensor decomposition.Experimental results show that the distributed and parallel method proposed in this paper can significantly improve the computational efficiency of the Hierarchical Tucker tensor decomposition on the base of ensuring the accuracy.2、A distributed and parallel tensor ring decomposition method based on cloud computing is proposed.In this paper,according to the characteristics of iterative computing tasks of tensor ring decomposition,a distributed random sampling tensor ring decomposition task division is designed to make the inseparable computing tasks run in a distributed environment and solve the bottleneck of large-scale tensor data decomposition.In the iterative process,the core tensors of each mode are calculated in parallel and distributed to accelerate the calculation.By sampling sorting,the original tensor data can be read only once in each iteration,thus reducing I/O consumption.The parallel optimization method of tensor ring decomposition based on cloud computing designed in this paper can effectively accelerate the operation of tensor decomposition,improve the ability to process tensor data in cloud computing,and support the application of tensor data on cloud platform.3、In order to apply tensor data on cloud platform better,this paper designs a tensor data resource management mode based on cloud computing.According to the application scenario of tensor data in cloud platform,this paper designs a tensor data resource management platform.Faced with the problems of long calculation time and high usage frequency of tensor decomposition,the tensor data is pre-decomposed and stored,which significantly improves the efficiency of users in using tensors on the cloud platform.