ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2015, Vol. 52 ›› Issue (4): 960-971.doi: 10.7544/issn1000-1239.2015.20131343

• 图形图像 • 上一篇    下一篇


冷芳玲1, 刘金鹏1, 王志刚1, 陈昌宁1, 鲍玉斌1, 于戈1, 邓超2   

  1. 1(东北大学信息科学与工程学院 沈阳 110819); 2(中国移动通信研究院业务支撑研究所 北京 100053) (
  • 出版日期: 2015-04-01
  • 基金资助: 
    基金项目:国家自然科学基金重点项目(61033007); 国家自然科学基金项目(61173028, 61272179);中央高校基本科研业务费专项基金项目(N100704001); 教育部-中国移动科研基金项目(MCM20125021)

Edge Cluster Based Large Graph Partitioning and Iterative Processing in BSP

Leng Fangling1,Liu Jinpeng1,Wang Zhigang1,Chen Changning1,Bao Yubin1,Yu Ge1,Deng Chao2   

  1. 1(College of Information Science and Engineering, Northeastern University, Shenyang 110819); 2(Division for Business Support, China Mobile Institute, Beijing 100053)
  • Online: 2015-04-01

摘要: 近年来随着互联网的普及和相关技术的日益成熟,大规模图数据处理成为新的研究热点.由于传统的如Hadoop等通用云平台不适合迭代式地处理图数据,研究人员基于BSP模型提出了新的处理方案,如Pregel,Hama,Giraph等.然而,图处理算法需要按照图的拓扑结构频繁交换中间计算结果而导致巨大的通信开销,这严重地影响了基于BSP模型的系统的处理性能.首先从降低消息通信的角度分析当前主流BSP系统的处理方案,然后提出了一种基于边聚簇的垂直混合划分策略(EC-VHP),并建立代价收益模型分析其消息通信优化的效果.在EC-VHP的基础上,提出了一个点-边计算模型,并设计了简单Hash索引和多队列并行顺序索引机制,进一步提高消息通信的处理效率.最后,在真实数据集和模拟数据集上的大量实验,验证了EC-VHP策略和索引机制的正确性和有效性.

关键词: 大规模图, BSP模型, 图划分, 点-边计算模型, 索引结构

Abstract: With the development of Internet and the gradual maturity of related techniques in recent years, the processing of large graphs has become a new hot research topic. Since it is not appropriate for traditional cloud computing platforms to process graph data iteratively, such as Hadoop, researchers have proposed some solutions based on the BSP model, such as Pregel, Hama and Giraph. However, since graph algorithms need to frequently exchange intermediate results in accordance with the graph’s topological structure, the tremendous communication overhead impacts the processing performance of systems based on the BSP model greatly. In this paper, we first analyze the solutions proposed by the well-known BSP-based systems in reducing communication overhead, and then propose a graph partition strategy named edge cluster based vertically hybrid partitioning (EC-VHP), building a cost benefit model to study its effectiveness to the communication overhead. Then based on EC-VHP, we propose a vertex-edge computation model, and design both a plain hash index structure and a multi-queue parallel sequential index structure to further improve the processing efficiency of message communication. Finally, our experiments on real and synthetic data sets demonstrate the efficiency and accuracy of the EC-VHP and the index mechanism.

Key words: large graph, BSP model, graph partition, vertex-edge computation model, index structure