文档章节

阅读论文 --- GPS: A Graph Processing System∗

wdfnst
 wdfnst
发布于 2016/05/16 22:12
字数 886
阅读 17
收藏 0
点赞 2
评论 0

输入图片说明

输入图片说明

  • Highlight, page 1
    GPS is similar to Google’s proprietary Pregel system, with three new features: (1) an extended API to make global computations more easily expressed and more efficient; (2) a dynamic repartitioning scheme that re- assigns vertices to different workers during the computation, based on messaging patterns; and (3) an optimization that distributes adjacency lists of high-degree vertices across all compute nodes to improve performance.

  • Highlight, page 1
    The GPS API has an extension that enables efficient implementation of al- gorithms composed of one or more vertex-centric com- putations, combined with global computations.

  • Highlight, page 1
    GPS has an optimization called large adjacency list par- titioning (LALP), which partitions the adjacency lists of high-degree vertices across compute nodes, again to reduce communication.

  • Highlight, page 2
    Broadly

  • Underline, page 2
    1.4 Partitioning Experiments

  • Highlight, page 3
    A computation thread loops through the vertices in the worker and executes vertex.compute() on each active ver- tex. It maintains an outgoing message buffer for all work- ers in the cluster, including itself. When a buffer is full it is either given to MINA threads for sending over the network, or passed directly to the local message parser thread.

  • Highlight, page 3
    MINA threads send and receive message buffers, as well as simple coordination messages between the master and the worker. When a message buffer is received, it is passed to the message parser thread.

  • Highlight, page 3
    A message parser thread parses incoming message buffers into separate messages and enqueues them into the re- ceiving vertices’ message queues for the next superstep.

  • Highlight, page 3
    GPS is implemented in Java. The compute nodes run HDFS (Hadoop Distributed File System) [18], which is used to store persistent data such as the input graph and the checkpointing files.

  • Highlight, page 3
    each line starts with the ID of a vertex u, followed by the IDs of u’s outgoing neighbors.

  • Highlight, page 3
    GPS assigns the vertices of G to workers using the same simple round- robin scheme used by Pregel: vertex u is assigned to worker W(u mod k).

  • Highlight, page 3
    The master coordinates the computation by instructing work- ers to: (a) start parsing input files; (b) start a new super- step; (c) terminate computation; and (d) checkpoint their states for fault-tolerance.

  • Highlight, page 5
    by partitioning large graphs “intelligently” before computation begins, we can reduce total network I/O by up to 13.6x and run-time by up to 2.5x. The effects of partitioning depend on three factors: (1) the graph algorithm being executed; (2) the graph itself; and (3) the configuration of the worker tasks across compute nodes.

  • Highlight, page 5
    Domain-based: In this partitioning scheme for web graphs only, we locate all web pages from the same domain in the same partition, and partition the domains randomly across the workers.

  • Highlight, page 5
    Random: The default “mod” partitioning method

  • Box, page 6
    (a) PageRank on sk-2005-d (b) Different algorithms on uk-2007-u Figure 6: Network I/O, different partitioning schemes (a) PageRank (50 iter.) on sk-2005-d (b) Different algorithms on uk-2007-u Figure 7: Run-time, different partitioning schemes

  • Highlight, page 8
    There are three questions any dynamic repartitioning scheme must answer: (1) which vertices to reassign; (2) how and when to move the reassigned vertices to their new workers; (3) how to lo- cate the reassigned vertices.

  • Highlight, page 8
    One option is to reassign vertex u at worker Wi to a new worker Wj if u send/receives more message to/from Wj than to/from any other worker, and that number of messages is over some threshold.

  • Highlight, page 9
    An obvious option is for each worker to store an in-memory map consisting of <vertex-id, new-worker-id> pairs.

  • Highlight, page 10
    We have developed a compiler from the Green-Marl [19] domain-specific language for graph processing into GPS.

  • Highlight, page 10
    some al- gorithms become very complex when implemented in a vertex-centric fashion—a classic example is contructing or doing a reverse traversal on a BFS tree. Green-Marl’s high-level constructs can express some of these compu- tations very easily, e.g., lines 7 and 11 in Figure 13b.

  • Box, page 11
    1 2 3 4 5 6 7 8 9 10 11 12 13 14 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Procedure PageRank(G: Graph, e,d: Double, PR: Node Prop<Double>(G)) { Int i = 0; Double N = G.NumNodes(); G.PR = 1 / N; // Init PageRank Do { // Main iteration diff = 0.0; Foreach (t: G.Nodes) { Double val = (1−d) / N + d∗Sum(w: t.InNbrs){ w.PR / w.OutDegree()}; t.PR <= val @ t; diff +=|val−t.PR|;} i++; }While(i<30);} (a) PageRank Procedure bc approx(G:Graph, BC:Node Prop<Float>) { G.BC = 0; // Initialize BC as 0 per each node Node Prop<Float> sigma, delta; G.sigma = 0; Node s = G.PickRandom(); s. sigma = 1; InBFS(v: G.Nodes From s) { // BFS−order traversal // Summing over BFS parents v.sigma = Sum(w:v.UpNbrs) { w.sigma }; } InReverse { // Reverse−BFS order traversal v.delta = // Summing over BFS children Sum (w:v.DownNbrs) { v.sigma / w.sigma ∗ (1+ w.delta) }; v.BC += v.delta; // accumulate delta into BC } (b) Approximate Betweenness Centrality Figure 13: Green-Marl Programs }

© 著作权归作者所有

共有 人打赏支持
wdfnst
粉丝 2
博文 18
码字总数 22859
作品 1
宁波
半定规划(SDP)for 最大割

半定规划(SDP)for 最大割 最大割问题: Instance: given an undirected graph G(V,E), find a bipartition of V into S and T that maximizing the size of the cut E(S,T)= 贪心算法、随机......

hlveying ⋅ 01/02 ⋅ 0

caffe学习:卷积计算

在caffe中如何计算卷积的? caffe中, 卷积网络的前向传播过程需要计算类似 这样的连接, forward_cpu_gemm()函数用weight矩阵和输入的bottom相乘, 然后对bias进行处理, bias程序会根据情况决定...

Hanging_Gardens ⋅ 01/11 ⋅ 0

100 open source Big Data architecture papers

Big Data technology has been extremely disruptive with open source playing a dominant role in shaping its evolution. While on one hand it has been disruptive, on the other it ha......

naughty ⋅ 2016/04/05 ⋅ 0

每周论文清单:知识图谱,文本匹配,图像翻译,视频对象分割

在碎片化阅读充斥眼球的时代,越来越少的人会去关注每篇论文背后的探索和思考。 在这个栏目里,你会快速 get 每篇精选论文的亮点和痛点,时刻紧跟 AI 前沿成果。 点击本文底部的「阅读原文」...

c9yv2cf9i06k2a9e ⋅ 2017/12/27 ⋅ 0

一种基于ASEF算法的高速特征点定位算法

一种基于ASEF算法的高速特征点定位算法 [TOC] 作者:侯伟 时间:2018.04.14 论文链接1:http://vision.lbl.gov/Conferences/cvpr/Papers/data/papers/1105.pdf 论文链接2:http://citeseerx....

monkey3233 ⋅ 04/13 ⋅ 0

Spark GraphX Pregel API: An Example

GraphX Pregel API Graph data and graph processing is getting more and more attention lately in various fields. It has become apparent that a large number of real world problems ......

openthings ⋅ 2016/09/02 ⋅ 1

『 Spark 』1. spark 简介

原文链接:『 Spark 』1. spark 简介 写在前面 本系列是综合了自己在学习spark过程中的理解记录 + 对参考文章中的一些理解 + 个人实践spark过程中的一些心得而来。写这样一个系列仅仅是为了...

litaotao ⋅ 2016/04/16 ⋅ 0

何恺明大神的「Focal Loss」,如何更好地理解?

作者丨苏剑林 单位丨广州火焰信息科技有限公司 研究方向丨NLP,神经网络 个人主页丨kexue.fm 前言 今天在 QQ 群里的讨论中看到了 Focal Loss,经搜索它是 Kaiming 大神团队在他们的论文 Foca...

c9yv2cf9i06k2a9e ⋅ 2017/12/28 ⋅ 0

实时计算框架 Spark: Lightning-fast cluster computing

real-time processing framework 实时计算框架 Real-time processing denotes processing, transforming and analyzing data on the fly Spark: Apache Spark has an advanced DAG executio......

swingcoder ⋅ 2016/08/24 ⋅ 0

『Big data technologies』关于各类大数据技术概念的简介(翻译自Quora)

原文 I'll try to give a very crude overview of how the pieces fit in together, because the details span multiple books. Please forgive me for some oversimplifications. MapReduce......

灰大羊 ⋅ 2016/07/18 ⋅ 0

没有更多内容

加载失败,请刷新页面

加载更多

下一页

Spring发展历程总结

转自与 https://www.cnblogs.com/RunForLove/p/4641672.html 目前很多公司的架构,从Struts2迁移到了SpringMVC。你有想过为什么不使用Servlet+JSP来构建Java web项目,而是采用SpringMVC呢?...

onedotdot ⋅ 41分钟前 ⋅ 0

Python模块/包/库安装(6种方法)

Python模块/包/库安装(6种方法) 冰颖机器人 2016-11-29 21:33:26 一、方法1: 单文件模块 直接把文件拷贝到 $python_dir/Lib 二、方法2: 多文件模块,带setup.py 下载模块包(压缩文件zip...

cswangyx ⋅ 今天 ⋅ 0

零基础学习大数据人工智能,学习路线篇!系统规划大数据之路?

大数据处理技术怎么学习呢?首先我们要学习Python语言和Linux操作系统,这两个是学习大数据的基础,学习的顺序不分前后。 Python:Python 的排名从去年开始就借助人工智能持续上升,现在它已经...

董黎明 ⋅ 今天 ⋅ 0

openJdk和sun jdk的区别

使用过LINUX的人都应该知道,在大多数LINUX发行版本里,内置或者通过软件源安装JDK的话,都是安装的OpenJDK, 那么到底什么是OpenJDK,它与SUN JDK有什么关系和区别呢? 历史上的原因是,Ope...

jason_kiss ⋅ 今天 ⋅ 0

梳理

Redux 是 JavaScript 状态容器,提供可预测化的状态管理。 它是JS的状态容器,是一种解决问题的方式,所以即可以用于 react 也可以用于 vue。 需要理解其思想及实现方式。 应用中所有的 stat...

分秒 ⋅ 今天 ⋅ 0

Java 后台判断是否为ajax请求

/** * 是否是Ajax请求 * @param request * @return */public static boolean isAjax(ServletRequest request){return "XMLHttpRequest".equalsIgnoreCase(((HttpServletReques......

JavaSon712 ⋅ 今天 ⋅ 0

Redis 单线程 为何却需要事务处理并发问题

Redis是单线程处理,也就是命令会顺序执行。那么为什么会存在并发问题呢? 个人理解是,虽然redis是单线程,但是可以同时有多个客户端访问,每个客户端会有 一个线程。客户端访问之间存在竞争...

码代码的小司机 ⋅ 今天 ⋅ 0

到底会改名吗?微软GVFS 改名之争

微软去年透露了 Git Virtual File System(GVFS)项目,GVFS 是 Git 版本控制系统的一个开源插件,允许 Git 处理 TB 规模的代码库,比如 270 GB 的 Windows 代码库。该项目公布之初就引发了争...

linux-tao ⋅ 今天 ⋅ 0

笔试题之Java基础部分【简】【二】

1.静态变量和实例变量的区别 在语法定义上的区别:静态变量前要加static关键字,而实例变量前则不加。在程序运行时的区别:实例变量属于某个对象的属性,必须创建了实例对象,其中的实例变...

anlve ⋅ 今天 ⋅ 0

Lombok简单介绍及使用

官网 通过简单注解来精简代码达到消除冗长代码的目的 优点 提高编程效率 使代码更简洁 消除冗长代码 避免修改字段名字时忘记修改方法名 4.idea中安装lombnok pom.xml引入 <dependency> <grou...

to_ln ⋅ 今天 ⋅ 0

没有更多内容

加载失败,请刷新页面

加载更多

下一页

返回顶部
顶部