基于CPU-Phi异构架构的高性能图计算系统

发布时间：2018-01-08 14:04

本文关键词：基于CPU-Phi异构架构的高性能图计算系统　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着大数据时代的到来,与大数据相关的技术变革成为学术界和工业界的关注热点。由于图在描述对象之间的联系方面具有丰富的表达能力,因此图被广泛的应用在交通路线的规划、论文著作的引用关系、安全事件分析、社交网络分析、生物医疗等许多领域。随着图数据的规模越来越大,分布式大图处理技术得到迅速发展,这些技术包括分布式图计算模型、大图数据存储与管理、各种高性能的图算法、高性能图计算系统等。Intel Xeon Phi是Intel面向高性能计算领域推出的一款协处理器。Phi在保持x86架构的基础上扩展了向量计算能力,具有计算能力强和易于开发的优点。Cgraph是目前唯一在CPU-Phi异构系统上实现的图计算系统,但是Cgraph在系统实现方面存在一些不足,如没有充分地利用Phi的向量化计算能力、消息缓冲机制低效、没有进行有效的线程管理等,另外Cgraph也没有实现对图数据的在线分割。在图分割方面,目前大多数图计算系统在划分图数据时未考虑计算节点的能力差异,从而不能充分利用异构计算节点的计算能力。本论文研究在CPU-Phi异构平台上实现高性能的图计算系统,对于充分的利用Phi体系结构特点去优化图计算做了有益的探索。论文主要工作包括以下三个方面:(1)充分利用Phi的体系结构特点优化了图计算过程;(2)针对CPU和Phi处理能力的不同设计了架构感知的图分割算法;(3)基于以上技术,在CPU-Phi异构架构上实现了一个单机版图计算系统Pgraph,并和Cgraph进行了实验比较。同时通过实验验证了 Pgraph在设计和实现方面的有效性。
[Abstract]:With the arrival of big data era, the technological changes related to big data have become the focus of attention in academia and industry. Therefore, maps are widely used in many fields, such as transportation route planning, citation relation of papers, security incident analysis, social network analysis, biomedicine and so on. With the increasing scale of map data. Distributed large graph processing technology has been developed rapidly. These technologies include distributed graph computing model, large graph data storage and management, and various high performance graph algorithms. High performance graph computing system etc. Intel Xeon. Phi is a coprocessor developed by Intel for high performance computing. Phi extends vector computing power while maintaining the x86 architecture. Cgraph is the only graph computing system implemented on CPU-Phi heterogeneous system. However, there are some shortcomings in the system implementation of Cgraph, such as not making full use of the vectorization computing ability of Phi, inefficient message buffering mechanism, and lack of effective thread management and so on. In addition, Cgraph does not realize the on-line segmentation of graph data. In the aspect of graph segmentation, most graph computing systems do not consider the difference of computing nodes' ability in dividing graph data. Therefore, we can not make full use of the computing power of heterogeneous computing nodes. This paper studies the implementation of high performance graph computing system on CPU-Phi heterogeneous platform. This paper makes a useful exploration to make full use of the characteristics of Phi architecture to optimize the graph calculation. The main work of this paper includes the following three aspects: 1). The graph computing process is optimized by making full use of the architecture characteristics of Phi. Secondly, an architecture aware graph segmentation algorithm is designed for different processing abilities of CPU and Phi. Based on the above technology, a single machine layout computing system, Pgraph, is implemented on the CPU-Phi heterogeneous architecture. The experimental results are compared with that of Cgraph, and the effectiveness of Pgraph in design and implementation is verified by experiments.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP332;TP311.13

【相似文献】