The Performance Testing And Optimization Of The Hybrid Architecture Cluster Based On CPU And GPU

As the science and technology is keeping developing and the social is keeping moving forward, high-performance computing technology appeared, especially in scientific computing. As the scale of the problem expands, the needs of the large-scale computing has growed rapidly. With the recent study, people found that the increase speed of the single CPU is slower and slower, people try to rise the efficience of the program by enhancing the performance of the CPU or by increasing the number of the cluster nodes, but this method not only need a high cost of money, but in most of cases, for a given parallel program, just increase the number of the compute nodes can not achieve to improve the program execution efficiency. In addition, with the development of the GPU architecture and the related innovative development platform, GPU computing in science has been widely used. Now the key point is how to develop the powerful multi-core computing components to achieve the level of the peak performance. Although people have built some hardware architecture and software platform, but the gap between parallel computing hardware platforms and corresponding software environment is growing. Based on this situation, this paper has done a deeply research of the hybrid architecture of the platform features and technical aspects of program optimization.In addition, this paper combines the performance of CPU cluster environment and the theory of the parallel program analysis, in the code aspect, the network communication aspect and the operating system aspect, this paper highlights some test cases and some improved cases, and also got some good results. The value of this paper is to strengthen the value of the high-performanced GPU computing applications and to highlight the advantage of the strong computing skill of GPU. Also it has some value to improve the efficiency of the parallel implementatio.