The Design And Implementation Of CPU‐GPU Heterogeneous Parallel Computing System

With the upgrading of CPU and GPU hardware in terminal device, how to provide users with high-performance computing environments has become a research hotspot. In the era of green high-performance computing, CPU-GPU heterogeneous computing system can provide good versatility, performance and productivity, and has broad prospects for development. In this paper, the systems optimize the original heterogeneous parallel technology of CPU-GPU from the aspects of the division of tasks and data transmission based on existing desktop system CPU and GPU hardware resources, and design and implement a CPU-GPU heterogeneous parallel computing system. The contents and results of this study are the following aspects:(1) The system proposed solutions to optimize heterogeneous parallel computing, and built heterogeneous parallel computing architecture after analyzing and studying of the multi-core CPU, GPU architecture of desktop system, as well as the existing heterogeneous parallel CPU-GPU technology. Then, heterogeneous parallel computing system is improved by deploying the node based on the desktop systems.(2) In the aspect of task allocation, tasks are divided by computing system in accordance with the feasibility of tasks. According to computing resources and bandwidth conditions isomers node, starting from the parallel execution time of the task, the system optimize the proposed load balancing scheduling programs to achieve optimal allocation of resources.(3) On the storage resource, computing system adopts the distributed storage approach, within the desktop, CPU-GPU heterogeneous parallel and uses virtual storage unified way through the overall distributed storage and local shared storage to balance the overall performance and programming complexity.(4) The system uses the method of CUDA + MPI to optimize distributed parallel communication and realize asynchronous processing mode, as well as the direct transfer of data between the GPU-GPU; on the basis of the original node inherent heterogeneous CPU-GPU communication, optimizing the heterogeneous communication method. Through the methods of data storage in the aspect of global array mapping, it reduces data transfer times and saves computing time.In this paper, focusing on parallel execution efficiency and performance for communication transmission system, the author tested the heterogeneous parallel computing systems by building matching database. By comparing the experimental data, optimization design of this article does significantly improve the computational performance compared to traditional CPU-GPU heterogeneous parallel computing architecture; the performance of heterogeneous parallel computing system is also superior to the traditional single GPU and CPU-led single CPU parallel computing system.