通用GPU计算技术在高性能计算平台上的应用研究

发布时间：2018-01-16 04:33

本文关键词：通用GPU计算技术在高性能计算平台上的应用研究　出处：《兰州大学》2012年硕士论文　论文类型：学位论文

【摘要】：近年来,随着芯片技术及生产工艺的不断提升,GPU凭借其高效的计算能力逐步进入高性能计算领域的研究。随着GPU计算能力和可编程性的不断提高,利用GPU进行通用计算(GPGPU)逐渐成为研究的热点。通常GPGPU计算采用GPU+CPU的混合架构模式,而GPU+CPU混合架构计算集群在硬件架构搭建、软件环境设计等方面与传统计算环境相比,又存在着明显的差异。对于承担大规模计算任务的计算中心来说,如何使得GPU计算技术高效的融入自身的计算研究工作已经成为一个十分必要的研究课题。本文基于甘肃省计算中心引入的混合架构集群进行应用研究工作,这种混合架构集群利用NVIDIA和ATI两类此领域主流的GPU搭建而成。研究内容主要涉及混合架构集群硬件架构方法、集群批作业自动处理、已有应用领域计算效率分析对比、GPU渲染应用等。研究包括了系统升级方案设计、计算效率测试、对比总结等工作。本论文所做的研究工作,在为甘肃省计算中心GPU高性能计算现有业务应用提出一套解决方案的同时,也为其今后设计、搭建更大规模的GPU+CPU混合架构集群,支撑省内大规模计算研究提供了一定的参考价值和技术支撑。
[Abstract]:In recent years, with the continuous improvement of chip technology and production technology, GPUs have gradually entered the field of high-performance computing by virtue of their efficient computing power, and with the continuous improvement of GPU computing power and programmability. Using GPU to carry out general purpose computing (GPU) has gradually become a hot research topic. Usually, GPGPU computing adopts the hybrid architecture mode of GPU CPU. The GPU CPU hybrid architecture computing cluster is compared with the traditional computing environment in hardware architecture, software environment design and so on. There is also a clear difference. For computing centers that undertake large-scale computing tasks. How to integrate GPU computing technology into its own computational research work has become a very necessary research topic. This paper is based on the hybrid architecture cluster introduced by Gansu Computing Center for application research. This hybrid architecture cluster is based on NVIDIA and ATI, which are two kinds of mainstream GPU in this field. The research mainly involves the hybrid architecture cluster hardware architecture method, cluster batch processing automatically. The research includes system upgrade scheme design, calculation efficiency test, comparison and summary. The research work done in this paper. This paper presents a solution for the existing GPU high-performance computing applications in Gansu Computing Center, and designs a larger GPU CPU hybrid architecture cluster for its future design. The research of large-scale calculation in support province provides certain reference value and technical support.
【学位授予单位】：兰州大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP338;TP391.41

【参考文献】