启动zookeeper_Giraph源码分析（一）—启动ZooKeeper服务

作者 | 白松

Giraph介绍：

Apache Giraph is an iterative graph processing system built for high scalability. For example, it is currently used at Facebook to analyze the social graph formed by users and their connections. Giraph originated as the open-source counterpart to Pregel, the graph processing architecture developed at Google and described in a 2010 paper. Both systems are inspired by the Bulk Synchronous Parallelmodel of distributed computation introduced by Leslie Valiant. Giraph adds several features beyond the basic Pregel model, including master computation, sharded aggregators, edge-oriented input, out-of-core computation, and more. With a steady development cycle and a growing community of users worldwide, Giraph is a natural choice for unleashing the potential of structured datasets at a massive scale.

原理：

Giraph基于Hadoop而建，将MapReduce中Mapper进行封装，未使用reducer。在Mapper中进行多次迭代，每次迭代等价于BSP模型中的SuperStep。一个Hadoop Job等价于一次BSP作业。基础结构如下图所示。

每部分的功能如下：

1. ZooKeeper: responsible for computation state

–partition/worker mapping

–global state: #superstep

–checkpoint paths, aggregator values, statistics

2. Master: responsible for coordination

–assigns partitions to workers

–coordinates synchronization

–requests checkpoints

–aggregates aggregator values

–collects health statuses

3. Worker: responsible for vertices

–invokes active vertices compute() function

–sends, receives and assigns messages

–computes local aggregation values

说明

(1)实验环境

三台服务器：test165、test62、test63。test165同时是JobTracker和TaskTracker.

测试例子：官网自带的SSSP程序，数据是自己模拟生成。

运行命令：Hadoop jar giraph-examples-1.0.0-for-hadoop-0.20.203.0-jar-with-dependencies.jar org.apache.giraph.GiraphRunner org.apache.giraph.examples.SimpleShortestPathsVertex -vif org.apache.giraph.io.formats.JsonLongDoubleFloatDoubleVertexInputFormat -vip /user/giraph/SSSP -of org.apache.giraph.io.formats.IdWithValueTextOutputFormat -op /user/giraph/output-sssp-debug-7 -w 5

(2)为节约空间，下文中所有代码均为核心代码片段。

(3)core-site.xml中hadoop.tmp.dir的路径设为：/home/hadoop/hadooptmp

(4)写本文是多次调试完成的，故文中的JobID不一样，读者可理解为同一JobID.

(5)后续文章也遵循上述规则。

org.apache.giraph.graph.GraphMapper类

Giraph中自定义org.apache.giraph.graph.GraphMapper类来继承Hadoop中的 org.apache.hadoop.mapreduce.Mapper类，覆写了setup()、map()、cleanup()和run()方法。GraphMapper类的说明如下：

“This mapper that will execute the BSP graph tasks alloted to this worker. All tasks will be performed by calling the GraphTaskManager object managed by this GraphMapper wrapper classs. Since this mapper will not be passing data by key-value pairs through the MR framework, the Mapper parameter types are irrelevant, and set to Object type.”

BSP的运算逻辑被封装在GraphMapper类中，其拥有一GraphTaskManager对象，用来管理Job的tasks。每个GraphMapper对象都相当于BSP中的一个计算节点(compute node)。

在GraphMapper类中的setup()方法中，创建GraphTaskManager对象并调用其setup()方法进行一些初始化工作。如下：

map()方法为空，因为所有操作都被封装在了GraphTaskManager类中。在run()方法中调用GraphTaskManager对象的execute()方法进行BSP迭代计算。

org.apache.giraph.graph.GraphMapper类

功能：The Giraph-specific business logic for a single BSP compute node in whatever underlying type of cluster our Giraph job will run on. Owning object will provide the glue into the underlying cluster framework and will call this object to perform Giraph work.

下面讲述setup()方法，代码如下：

依次介绍每个方法的功能：

1、locateZookeeperClasspath(zkPathList)

找到ZK jar的本地副本，其路径为：/home/hadoop/hadooptmp/mapred/local/taskTracker/root/jobcache/job_201403270456_0001/jars/job.jar ,用于启动ZooKeeper服务。

2、startZooKeeperManager()，初始化和配置ZooKeeperManager。

定义如下：

3、org.apache.giraph.zk.ZooKeeperManager 类

功能：Manages the election of ZooKeeper servers, starting/stopping the services, etc.

ZooKeeperManager类的setup()定义如下：

createCandidateStamp()方法在 HDFS上的_bsp/_defaultZkManagerDir/job_201403301409_0006/_task 目录下为每个task创建一个文件，文件内容为空。文件名为本机的Hostname+taskPartition，如下截图：

运行时指定了5个workers(-w 5)，再加上一个master，所有上面有6个task。

getZooKeeperServerList()方法中，taskPartition为0的task会调用createZooKeeperServerList()方法创建ZooKeeper server List，也是创建一个空文件，通过文件名来描述Zookeeper servers。

首先获取taskDirectory(_bsp/_defaultZkManagerDir/job_201403301409_0006/_task)目录下文件，如果当前目录下有文件，则把文件名(Hostname+taskPartition)中的Hostname和taskPartition存入到hostNameTaskMap中。扫描taskDirectory目录后，若hostNameTaskMap的size大于serverCount(等于GiraphConstants.java中的ZOOKEEPER_SERVER_COUNT变量，定义为1)，就停止外层的循环。外层循环的目的是：因为taskDirectory下的文件每个task文件时多个task在分布式条件下创建的，有可能task 0在此创建server List时，别的task还没有生成后task文件。Giraph默认为每个Job启动一个ZooKeeper服务，也就是说只有一个task会启动ZooKeeper服务。

经过多次测试，task 0总是被选为ZooKeeper Server ，因为在同一进程中，扫描taskDirectory时，只有它对应的task 文件(其他task的文件还没有生成好)，然后退出for循环，发现hostNameTaskMap的size等于1，直接退出while循环。那么此处就选了test162 0。

最后，创建了文件：_bsp/_defaultZkManagerDir/job_201403301409_0006/zkServerList_test162 0

onlineZooKeeperServers()，根据zkServerList_test162 0文件，Task 0 先生成zoo.cfg配置文件，使用ProcessBuilder来创建ZooKeeper服务进程，然后Task 0 再通过socket连接到ZooKeeper服务进程上，最后创建文件 _bsp/_defaultZkManagerDir/job_201403301409_0006/_zkServer/test162 0 来标记master任务已完成。worker一直在进行循环检测master是否生成好 _bsp/_defaultZkManagerDir/job_201403301409_0006/_zkServer/test162 0，即worker等待直到master上的ZooKeeper服务已经启动完成。

启动ZooKeeper服务的命令如下：

4、determineGraphFunctions()。

GraphTaskManager类中有CentralizedServiceMaster对象和CentralizedServiceWorker 对象，分别对应于master和worker。每个BSP compute node扮演的角色判定逻辑如下：

a) If not split master, everyone does the everything and/or running ZooKeeper.

b) If split master/worker, masters also run ZooKeeper

c) If split master/worker == true and giraph.zkList is set, the master will not instantiate a ZK instance, but will assume a quorum is already active on the cluster for Giraph to use.

该判定在GraphTaskManager 类中的静态方法determineGraphFunctions()中定义，片段代码如下：