| GPU has been used not only in the field of graphic images,in recent years because of its architectural features and the promote of the floating point developed of the whole numerical computing domain.Especially for easy parallel processing tasks,computing time can be more than several times or even 10 times of the speed.In many areas of large numerical computation,GPU has played a significant role indeed.In the field of the machine learning,logistic regression could reach hundreds of millions of features,so a simple stand-alone version of the training,even the parallel processing has been unable to meet the training requirements.The best solution is to split the high dimensional feature vectors into several small vectors.The emergence of parallel machine learning algorithms improved the current situation,thousands or even tens of thousands of machines train the data in parallel so as to improve the speed of execution.In the energy industry because of the low-carbon,high energy density,high duration and other characteristics,the development trend of nuclear energy can not be replaced easily by other energy sources.The management of the core fuel is one of the major problems to be concerned by the nuclear power plant,and it is directly related to the economic benefits and costs of nuclear power.The order number of large-scale reactor core expansion equation is large and complicated,so the calculation of the diffusion equation is crucial in core fuel management process.The main research work and contributions of this paper are as follows:(1)In this paper,we use CUDA programming technology on the GPU cluster to run the LU-SGS parallel iterative calculation,set against background of the unstructured grid and fluid field.Divide one-dimensional and two-dimensional mesh into a plurality of domains,each domain is evenly distributed on the GPU blocks.And put forward a parallel collision scheme that the virtual boundary of the grid exchange information and data transmission by non-stop collision.Compared with the serial program,the experiment shows that GPU greatly improves the efficiency of the program execution,and verify that GPU is playing a much more important role in the field of numerical calculations.(2)This paper derives and analyzes the parallel logistic regression algorithm,proposes the theory of using Raft algorithm instead of DHT algorithm to change the consistency requirements of the parameter server,and analyzes the importance of using GPU on the parameters server.(3)This paper LU-SGS iterative method and the parallel logistic regression algorithm,they are not parallel high degree of algorithm.Since the threads between the GPU bottom blocks can not communicate with each other,the shared memory can not access each other.Therefore,it is not possible to calculate the data by "collision".This paper proposes a method of "delay" iteration,Or iterates one step backward.This method can reduce the execution time of LU-SGS algorithm by about 20%. |
