Research And Application Of High Performance Computing Based On GPU

In recent years, the rapid development and popularization of information technology have brought about the explosive growth of data, which presents new challenges to the existing data processing technology. The scale of big data is so large that makes it difficult for serial computation method to quickly complete their processing and computing within an acceptable time. In order to improve efficiency, HPC technology is spontaneously required to support big data processing by divide-and-conquer parallel computing model. On the other hand, GPU has become the mainstream of HPC accelerator by its powerful parallel computing capability, high throughput and cost-effective.However, the utilization of HPC capacity of GPU is insufficient in current GPU-based HPC technology which making it difficult to deal with the parallel computing in big data environment. In addition, some application scenarios require users to be familiar with the details of GPU program development, which reduce the ease of use of high performance computing system.Therefore, this paper focuses on two aspects of GPU-based HPC under big data environment:(1) Research and improvement on parallel computing model of GPU-based HPC:Map Reduce is a distributed computing model which is suitable for big data processing, but its computing power is limited by hardware devices. Therefore, this paper designs and implements a Map Reduce-based GPU parallel computing model——GSMR, on the basis of Map Reduce model and the aid of the strong hardware parallelism of GPU. The experiment results show that the speedup of GSMR compared with the similar model is well, and the model is scalable.(2) Research on specific applications of GPU HPC technology, in particular in two issues:Research the problem of throughput rate shortage when traditional network device handles IP packets under big data environment. This paper presents a GPU-based parallel packet classification method, analysis its parallelization and optimization method. The experiments show that this mothed can effectively improve the packet processing speed and the throughput rate of network device.Research that how to provide GPU HPC virtualization services regarding to scientific computing users. Thus, a GPU virtualization method based RPC is proposed. By way of function-level calls, there is no need for users to care about GPU programming details. The experiments show that this method greatly improve the speed relative to the local computing and enhance the usability of HPC.