基于GPGPU技术的大规模地理数据的处理和分析

发布时间：2018-05-02 16:29

本文选题：GPU + CUDA　；参考：《中南大学》2013年硕士论文

【摘要】：摘要：随着地理信息系统技术、遥感技术、计算机技术以及信息技术的快速发展,以多种技术结合为手段,从海量的实时数据源中提取有用信息,用以解决资源管理配置、城市规划管理、土地信息管理、生态环境管理、基础设施建设、交通规划等问题。然而,对于海量数据的处理,其运算量相对于一般的处理来说会有几百到几千倍的增加,不仅严重考验着计算机的数据处理能力,而且考验着算法设计在处理海量数据问题中的有效性。高性能计算技术突飞猛进的发展,给海量数据的处理工作带来了新的方向。多核CPU技术以及图形处理器(GPU)日益增强的可编程性以及高效计算能力,促进处理方式的巨大变化,由以往的CPU端编程处理,逐渐过渡到CPU+GPU异构编程处理,再发展到分布式处理,以及云计算。处理方式的改变带动了处理效率的提高,并由以往单一的计算机到现今多台计算机同时计算,同时实现了事务处理的均衡分配,可有效的利用各种计算机资源以提高处理效率。本文中结合GPU并行计算技术,探讨如何利用GPU存储器特点完成对海量数据的处理任务,以取得较好的加速效果。本文所做的工作有如下几个方面： 1、针对海量遥感影像数据的切分与调度问题,提出了基于CPU+GPU异构编程快速处理影像数据的解决方案,探讨利用不同的GPU存储器实现遥感影像数据的重采样处理。基于GPU并行技术的影像处理,应考虑到算法的设计和处理任务的划分,合理的划分线程,实现并行执行优化、存储器优化、指令使用优化,以提高整体的处理效率。阐述了统一计算设备架构(Compute Unified Device Architecture, CUDA)通用计算模型构架及其特点,并在此基础上实现了对于遥感影像数据的重采样加速。 2、提出GPU并行处理空间聚类中复杂的数值计算问题。以二部图空间聚类算法为例,依据聚类中数值计算的特点,以及GPU并行结构和硬件特点,探讨合适的并行处理方式,采用全局存储器、共享存储器加速技术,提高了数据的处理效率。实验结果表明,基于GPU并行计算比CPU串行计算在效率上有显著的提高。图14幅,表7个,参考文献47篇。
[Abstract]:Abstract: with the rapid development of geographic information system technology, remote sensing technology, computer technology and information technology, useful information is extracted from massive real-time data sources to solve the problem of resource management. Urban planning management, land information management, ecological environment management, infrastructure construction, traffic planning and other issues. However, for the processing of massive data, its computation will increase hundreds to thousands times compared with the normal processing, which not only seriously tests the computer's data processing ability, It also tests the effectiveness of algorithm design in dealing with massive data problems. The rapid development of high performance computing technology has brought a new direction to the processing of massive data. The increasing programmability and high efficiency computing power of multi-core CPU technology and graphics processor (GPU) promote the great change of processing methods, and gradually transition from the former CPU side programming processing to the CPU GPU heterogeneous programming processing. And then into distributed processing, and cloud computing. The change of processing mode leads to the improvement of processing efficiency, and from the former single computer to many computers at the same time, it realizes the balanced allocation of transaction processing, which can effectively use all kinds of computer resources to improve the processing efficiency. Combined with GPU parallel computing technology, this paper discusses how to use the characteristics of GPU memory to complete the processing of massive data in order to achieve a better acceleration effect. The work done in this paper is as follows: 1. Aiming at the problem of segmentation and scheduling of massive remote sensing image data, a solution of fast processing of remote sensing image data by heterogeneous programming based on CPU GPU is proposed, and different GPU memory is used to realize resampling processing of remote sensing image data. The image processing based on GPU parallel technology should consider the design of algorithm and partition of processing tasks, reasonably partition threads, realize parallel execution optimization, memory optimization, instruction usage optimization, in order to improve the overall processing efficiency. In this paper, the general computing model architecture of computer Unified Device Architecture (CUDAA) and its characteristics are described. On this basis, the resampling acceleration of remote sensing image data is realized. 2. GPU parallel processing of complex numerical computation problem in spatial clustering is proposed. Taking the bipartite graph space clustering algorithm as an example, according to the characteristics of numerical calculation in clustering, as well as the parallel structure and hardware characteristics of GPU, the suitable parallel processing method is discussed. The global memory and shared memory acceleration technology are adopted. The efficiency of data processing is improved. Experimental results show that the efficiency of parallel computing based on GPU is significantly higher than that of serial computing based on CPU. There are 14 figures, 7 tables and 47 references.
【学位授予单位】：中南大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP751;P208

【参考文献】