Implemention,Optimization And Application Of CUDA-based R Package For Parallel Acceleration

In recent years,with the development of Internet technology,the information began exploding.Data mining technology is getting more and more attention from the public as it can extract valuable information from massive data.R,as an interpretive programming language,has become a powerful open software for statistical science and data analysis because of its concise and flexible programming method as well as powerful third-party toolkit.However,due to the inherent problem of low computational efficiency,timeliness becomes a bottleneck of data analysis in R when faced with huge quantity of calculation.Therefore,it is very important to improve the performance of R code for promoting further development of data analysis technology using R.On the other hand,GPU has been successfully used in many parallel accelerated applications owing to its powerful ability in parallel processing,flexible programming and low cost,etc.Therefore,the GPU-accelerated parallel calculation has become a trend of improving the performance of R code,and has great research and application prospect.Based on the application status of GPU-accelerated parallel computing in R code optimization,we research and implement a R package accelerated by GPU.On this basis,we transplant and optimize deep learning algorithm based on GPU in order to speed up deep network training.In this paper,the main achievements are as follows:1.In this paper,we design and implement a R package using CUDA according to R user's habit.This R package can support Windows and NUIX operating system,and implement automatic system environment's detection and configuration while installing and importing.It has good compatibility and usability.This paper construct a hierarchical system architecture by using object-oriented programming method,which enables transparent use of GPU computing functions and flexible switch between the runtime device(CPU/GPU).This system has good flexibility,robustness and scalability.In addition,we optimize the storage and reuse problem of intermediate results based on the system architecture so that the efficiency of program is improved.2.This paper do exploration on the parallel optimization of deep learning algorithm for R and implement DBN training based on CPU +GPU heterogeneous programming model.This method is flexible and the speed up effect is striking.