Design And Implementation Of Parallel SVM Based On GPU

With the expansion of computing needs, and the development of information technology and bio-genetic technologies, high-performance computing has become one of the most popular field of study. The current level of development of high-performance computing is not only a key indicators of a country's comprehensive national strength and international competitiveness, and also the strategic high ground that every country competes for.A dozen of years have passed since the first emergence of GPU. In a very long time GPU can only be used for video encoding and decoding,3D rendering and other image processing fields. With the continuous improvement of performance and enhancement of functionality, GPU has gradually been applied to scientific computing field which requires substantial computing power. People have started using GPU to accelerate the matrix multiplication since 2001, and from then on, an explosive growth occurred in the applications that use GPU as an accelerator. The use of GPU by scientific community has also changed the structure of it greatly, in order to meet the needs of scientific computing, the structure of GPU is gradually changed to better suit general purpose needs. It can not only speed up the graphics, but also accelerate scientific computing applications, high performance computing based on GPU has gradually become a research hot spots.There are lots of defects in applications of SVM with large-scale data, such as the slow training speed and the high memory resources consuming. It is very promising to design an appropriate algorithm and make use of GPU in order to improve the practicality of SVM.This paper first analyzes the parallel points in support vector machine training and prediction algorithms, and designs a parallel algorithm. Then with the further insight of the programming based on GPU, choose OpenCL as an implementation tool, and complete the development of program under Visual Studio 2010 IDE. Finally, simulations have been carried out on both Core i7 980X+GTX 470 platform and Q9400+Geforce 310 platform. The design of software and experimental results are introduced in detail. Experimental results show that the time used in training and prediction of support vector machine reduces greatly when parallel algorithm based on GPU is adopted. It is feasible and has some practical values to some degree.