hadoop2 作业执行过程之作业提交

hadoop2.2.0、centos6.5

hadoop任务的提交常用的两种，一种是测试常用的IDE远程提交，另一种就是生产上用的客户端命令行提交

通用的任务程序提交步骤为：

1.将程序打成jar包；

2.将jar包上传到HDFS上；

3.用命令行提交HDFS上的任务程序。

跟着提交步骤从命令行提交开始

最简单的提交命令应该如：

hadoop jar /home/hadoop/hadoop-2.2.0/hadoop-examples.jar wordcount inputPath outputPath

在名为hadoop的shell 命令文件中当参数为jar时

确定了要运行的CLASS文件和环境变量后最后执行了了exec命令来运行

看org.apache.hadoop.util.RunJar类的main方法

 1 public static void main(String[] args) throws Throwable {
 2     String usage = "RunJar jarFile [mainClass] args...";
 3     //验证提交的参数数量
 4     if (args.length < 1) {
 5       System.err.println(usage);
 6       System.exit(-1);
 7     }
 8     //验证jar文件是否存在
 9     int firstArg = 0;
10     String fileName = args[firstArg++];
11     File file = new File(fileName);
12     if (!file.exists() || !file.isFile()) {
13       System.err.println("Not a valid JAR: " + file.getCanonicalPath());
14       System.exit(-1);
15     }
16     String mainClassName = null;
17 
18     JarFile jarFile;
19     try {
20       jarFile = new JarFile(fileName);
21     } catch(IOException io) {
22       throw new IOException("Error opening job jar: " + fileName)
23         .initCause(io);
24     }
25     //验证是否存在main方法
26     Manifest manifest = jarFile.getManifest();
27     if (manifest != null) {
28       mainClassName = manifest.getMainAttributes().getValue("Main-Class");
29     }
30     jarFile.close();
31 
32     if (mainClassName == null) {
33       if (args.length < 2) {
34         System.err.println(usage);
35         System.exit(-1);
36       }
37       mainClassName = args[firstArg++];
38     }
39     mainClassName = mainClassName.replaceAll("/", ".");
40     //设置临时目录并验证
41     File tmpDir = new File(new Configuration().get("hadoop.tmp.dir"));
42     ensureDirectory(tmpDir);
43 
44     final File workDir;
45     try { 
46       workDir = File.createTempFile("hadoop-unjar", "", tmpDir);
47     } catch (IOException ioe) {
48       // If user has insufficient perms to write to tmpDir, default  
49       // "Permission denied" message doesn't specify a filename. 
50       System.err.println("Error creating temp dir in hadoop.tmp.dir "
51                          + tmpDir + " due to " + ioe.getMessage());
52       System.exit(-1);
53       return;
54     }
55 
56     if (!workDir.delete()) {
57       System.err.println("Delete failed for " + workDir);
58       System.exit(-1);
59     }
60     ensureDirectory(workDir);
61     //增加删除工作目录的钩子，任务执行完后要删除
62     ShutdownHookManager.get().addShutdownHook(
63       new Runnable() {
64         @Override
65         public void run() {
66           FileUtil.fullyDelete(workDir);
67         }
68       }, SHUTDOWN_HOOK_PRIORITY);
69 
70 
71     unJar(file, workDir);
72 
73     ArrayList<URL> classPath = new ArrayList<URL>();
74     classPath.add(new File(workDir+"/").toURI().toURL());
75     classPath.add(file.toURI().toURL());
76     classPath.add(new File(workDir, "classes/").toURI().toURL());
77     File[] libs = new File(workDir, "lib").listFiles();
78     if (libs != null) {
79       for (int i = 0; i < libs.length; i++) {
80         classPath.add(libs[i].toURI().toURL());
81       }
82     }
83     //通过反射的方式执行任务程序的main方法，并把剩余的参数作为任务程序main方法的参数
84     ClassLoader loader =
85       new URLClassLoader(classPath.toArray(new URL[0]));
86 
87     Thread.currentThread().setContextClassLoader(loader);
88     Class<?> mainClass = Class.forName(mainClassName, true, loader);
89     Method main = mainClass.getMethod("main", new Class[] {
90       Array.newInstance(String.class, 0).getClass()
91     });
92     String[] newArgs = Arrays.asList(args)
93       .subList(firstArg, args.length).toArray(new String[0]);
94     try {
95       main.invoke(null, new Object[] { newArgs });
96     } catch (InvocationTargetException e) {
97       throw e.getTargetException();
98     }
99   }

环境设置好后就要开始执行任务程序的main方法了
以WordCount为例：

 1 package org.apache.hadoop.examples;
 2 
 3 import java.io.IOException;
 4 import java.util.StringTokenizer;
 5 
 6 import org.apache.hadoop.conf.Configuration;
 7 import org.apache.hadoop.fs.Path;
 8 import org.apache.hadoop.io.IntWritable;
 9 import org.apache.hadoop.io.Text;
10 import org.apache.hadoop.mapreduce.Job;
11 import org.apache.hadoop.mapreduce.Mapper;
12 import org.apache.hadoop.mapreduce.Reducer;
13 import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
14 import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
15 import org.apache.hadoop.util.GenericOptionsParser;
16 
17 public class WordCount {
18 
19   public static class TokenizerMapper 
20        extends Mapper<Object, Text, Text, IntWritable>{
21     
22     private final static IntWritable one = new IntWritable(1);
23     private Text word = new Text();
24       
25     public void map(Object key, Text value, Context context
26                     ) throws IOException, InterruptedException {
27       StringTokenizer itr = new StringTokenizer(value.toString());
28       while (itr.hasMoreTokens()) {
29         word.set(itr.nextToken());
30         context.write(word, one);
31       }
32     }
33   }
34   
35   public static class IntSumReducer 
36        extends Reducer<Text,IntWritable,Text,IntWritable> {
37     private IntWritable result = new IntWritable();
38 
39     public void reduce(Text key, Iterable<IntWritable> values, 
40                        Context context
41                        ) throws IOException, InterruptedException {
42       int sum = 0;
43       for (IntWritable val : values) {
44         sum += val.get();
45       }
46       result.set(sum);
47       context.write(key, result);
48     }
49   }
50 
51   public static void main(String[] args) throws Exception {
52     Configuration conf = new Configuration();
53     String[] otherArgs = new GenericOptionsParser(conf, args).getRemainingArgs();
54     if (otherArgs.length != 2) {
55       System.err.println("Usage: wordcount <in> <out>");
56       System.exit(2);
57     }
58     Job job = new Job(conf, "word count");
59     job.setJarByClass(WordCount.class);
60     job.setMapperClass(TokenizerMapper.class);
61     job.setCombinerClass(IntSumReducer.class);
62     job.setReducerClass(IntSumReducer.class);
63     job.setOutputKeyClass(Text.class);
64     job.setOutputValueClass(IntWritable.class);
65     FileInputFormat.addInputPath(job, new Path(otherArgs[0]));
66     FileOutputFormat.setOutputPath(job, new Path(otherArgs[1]));
67     System.exit(job.waitForCompletion(true) ? 0 : 1);
68   }
69 }

在程序运行入口main方法中

首先定义配置文件类 Configuration，此类是Hadoop各个模块的公共使用类，用于加载类路径下的各种配置文件，读写其中的配置选项；

第二步中用到了 GenericOptionsParser 类，其目的是将命令行中的后部分参数自动设置到变量conf中，

如果代码提交的时候传入其他参数，比如指定reduce的个数，可以根据 GenericOptionsParser的命令行格式这么写：

bin/hadoop jar MyJob.jar com.xxx.MyJobDriver -Dmapred.reduce.tasks=5，

其规则是 -D 加上MR的配置选项（默认reduce task的个数为1，map的个数也为1）；

之后就是 Job 的定义

使用的job类的构造方法为

  public Job(Configuration conf, String jobName) throws IOException {
    this(conf);
    setJobName(jobName);
  }

调用了另外一个构造方法，并设置了Job的名字（即WordCount）

  public Job(Configuration conf) throws IOException {
    this(new JobConf(conf));
  }

  public JobConf(Configuration conf) {
    super(conf);
    
    if (conf instanceof JobConf) {
      JobConf that = (JobConf)conf;
      credentials = that.credentials;
    }
    
    checkAndWarnDeprecation();
  }

job 已经根据配置信息实例化好运行环境了，下面就是加入实体“口食”

依次给job添加Jar包、设置Mapper类、设置合并类、设置Reducer类、设置输出键类型、设置输出值类型

在setJarByClass中

  public void setJarByClass(Class<?> cls) {
    ensureState(JobState.DEFINE);
    conf.setJarByClass(cls);
  }

它先判断当前job的状态是否在运行中，接着通过class找到jar文件，将jar路径赋值给mapreduce.jar.jar属性（寻找jar文件的方法使通过ClassUtil类中的findContainingJar方法）

job的提交方法是

job.waitForCompletion(true)

 1   public boolean waitForCompletion(boolean verbose
 2                                    ) throws IOException, InterruptedException,
 3                                             ClassNotFoundException {
 4     if (state == JobState.DEFINE) {
 5       submit();
 6     }
 7     if (verbose) {
 8       monitorAndPrintJob();
 9     } else {
10       // get the completion poll interval from the client.
11       int completionPollIntervalMillis = 
12         Job.getCompletionPollInterval(cluster.getConf());
13       while (!isComplete()) {
14         try {
15           Thread.sleep(completionPollIntervalMillis);
16         } catch (InterruptedException ie) {
17         }
18       }
19     }
20     return isSuccessful();
21   }

参数 verbose ，如果想在控制台打印当前的任务执行进度，则设为true

 1   public void submit() 
 2          throws IOException, InterruptedException, ClassNotFoundException {
 3     ensureState(JobState.DEFINE);
 4     setUseNewAPI();
 5     connect();
 6     final JobSubmitter submitter = 
 7         getJobSubmitter(cluster.getFileSystem(), cluster.getClient());
 8     status = ugi.doAs(new PrivilegedExceptionAction<JobStatus>() {
 9       public JobStatus run() throws IOException, InterruptedException, 
10       ClassNotFoundException {
11         return submitter.submitJobInternal(Job.this, cluster);
12       }
13     });
14     state = JobState.RUNNING;
15     LOG.info("The url to track the job: " + getTrackingURL());
16    }

在submit 方法中会把Job提交给对应的Cluster，然后不等待Job执行结束就立刻返回

同时会把Job实例的状态设置为JobState.RUNNING，从而来表示Job正在进行中

然后在Job运行过程中，可以调用getJobState()来获取Job的运行状态

Submit主要进行如下操作

检查Job的输入输出是各项参数，获取配置信息和远程主机的地址，生成JobID，确定所需工作目录(也是MRAppMaster.java所在目录)，执行期间设置必要的信息
拷贝所需要的Jar文件和配置文件信息到HDFS系统上的指定工作目录，以便各个节点调用使用
计算并获数去输入分片(Input Split)的数目，以确定map的个数
调用YARNRunner类下的submitJob()函数，提交Job，传出相应的所需参数(例如 JobID等)。
等待submit()执行返回Job执行状态，最后删除相应的工作目录。

在提交前先链接集群（cluster），通过connect方法

 1   private synchronized void connect()
 2           throws IOException, InterruptedException, ClassNotFoundException {
 3     if (cluster == null) {
 4       cluster = 
 5         ugi.doAs(new PrivilegedExceptionAction<Cluster>() {
 6                    public Cluster run()
 7                           throws IOException, InterruptedException, 
 8                                  ClassNotFoundException {
 9                      return new Cluster(getConfiguration());
10                    }
11                  });
12     }
13   }

这是一个线程保护方法。这个方法中根据配置信息初始化了一个Cluster对象，即代表集群

 1   public Cluster(Configuration conf) throws IOException {
 2     this(null, conf);
 3   }
 4 
 5   public Cluster(InetSocketAddress jobTrackAddr, Configuration conf) 
 6       throws IOException {
 7     this.conf = conf;
 8     this.ugi = UserGroupInformation.getCurrentUser();
 9     initialize(jobTrackAddr, conf);
10   }
11   
12   private void initialize(InetSocketAddress jobTrackAddr, Configuration conf)
13       throws IOException {
14 
15     synchronized (frameworkLoader) {
16       for (ClientProtocolProvider provider : frameworkLoader) {
17         LOG.debug("Trying ClientProtocolProvider : "
18             + provider.getClass().getName());
19         ClientProtocol clientProtocol = null; 
20         try {
21           if (jobTrackAddr == null) {
　　　　　　　　  //创建YARNRunner对象  
22             clientProtocol = provider.create(conf);
23           } else {
24             clientProtocol = provider.create(jobTrackAddr, conf);
25           }
26           //初始化Cluster内部成员变量
27           if (clientProtocol != null) {
28             clientProtocolProvider = provider;
29             client = clientProtocol;
30             LOG.debug("Picked " + provider.getClass().getName()
31                 + " as the ClientProtocolProvider");
32             break;
33           }
34           else {
35             LOG.debug("Cannot pick " + provider.getClass().getName()
36                 + " as the ClientProtocolProvider - returned null protocol");
37           }
38         } 
39         catch (Exception e) {
40           LOG.info("Failed to use " + provider.getClass().getName()
41               + " due to error: " + e.getMessage());
42         }
43       }
44     }
45 
46     if (null == clientProtocolProvider || null == client) {
47       throw new IOException(
48           "Cannot initialize Cluster. Please check your configuration for "
49               + MRConfig.FRAMEWORK_NAME
50               + " and the correspond server addresses.");
51     }
52   }

可以看出创建客户端代理阶段使用了java.util.ServiceLoader，在2.3.0版本中包含LocalClientProtocolProvider（本地作业）和YarnClientProtocolProvider（yarn作业）(hadoop有一个Yarn参数mapreduce.framework.name用来控制你选择的应用框架。在MRv2里，mapreduce.framework.name有两个值：local和yarn)，此处会根据mapreduce.framework.name的配置创建相应的客户端

（ServiceLoader是服务加载类，它根据文件配置来在java classpath环境中加载对应接口的实现类）

这里在实际生产中一般都是yarn，所以会创建一个YARNRunner对象（客户端代理类）类进行任务的提交

实例化Cluster后开始真正的任务提交

submitter.submitJobInternal(Job.this, cluster)

  1 JobStatus submitJobInternal(Job job, Cluster cluster)   
  2 throws ClassNotFoundException, InterruptedException, IOException {  
  3   
  4   
  5   //检测输出目录合法性，是否已存在，或未设置  
  6   checkSpecs(job);  
  7   
  8   
  9   Configuration conf = job.getConfiguration();  
 10   addMRFrameworkToDistributedCache(conf);  
 11     //获得登录区，用以存放作业执行过程中用到的文件，默认位置/tmp/hadoop-yarn/staging/root/.staging ，可通过yarn.app.mapreduce.am.staging-dir修改  
 12   Path jobStagingArea = JobSubmissionFiles.getStagingDir(cluster, conf);  
 13   //主机名和地址设置  
 14   InetAddress ip = InetAddress.getLocalHost();  
 15   if (ip != null) {  
 16     submitHostAddress = ip.getHostAddress();  
 17     submitHostName = ip.getHostName();  
 18     conf.set(MRJobConfig.JOB_SUBMITHOST,submitHostName);  
 19     conf.set(MRJobConfig.JOB_SUBMITHOSTADDR,submitHostAddress);  
 20   }  
 21   //获取新的JobID，此处需要RPC调用  
 22   JobID jobId = submitClient.getNewJobID();  
 23   job.setJobID(jobId);  
 24   //获取提交目录：/tmp/hadoop-yarn/staging/root/.staging/job_1395778831382_0002  
 25   Path submitJobDir = new Path(jobStagingArea, jobId.toString());  
 26   JobStatus status = null;  
 27   try {  
 28     conf.set(MRJobConfig.USER_NAME,  
 29         UserGroupInformation.getCurrentUser().getShortUserName());  
 30     conf.set("hadoop.http.filter.initializers",   
 31         "org.apache.hadoop.yarn.server.webproxy.amfilter.AmFilterInitializer");  
 32     conf.set(MRJobConfig.MAPREDUCE_JOB_DIR, submitJobDir.toString());  
 33     LOG.debug("Configuring job " + jobId + " with " + submitJobDir   
 34         + " as the submit dir");  
 35     // get delegation token for the dir  
 36     TokenCache.obtainTokensForNamenodes(job.getCredentials(),  
 37         new Path[] { submitJobDir }, conf);  
 38       
 39     populateTokenCache(conf, job.getCredentials());  
 40   
 41   
 42     // generate a secret to authenticate shuffle transfers  
 43     if (TokenCache.getShuffleSecretKey(job.getCredentials()) == null) {  
 44       KeyGenerator keyGen;  
 45       try {  
 46         keyGen = KeyGenerator.getInstance(SHUFFLE_KEYGEN_ALGORITHM);  
 47         keyGen.init(SHUFFLE_KEY_LENGTH);  
 48       } catch (NoSuchAlgorithmException e) {  
 49         throw new IOException("Error generating shuffle secret key", e);  
 50       }  
 51       SecretKey shuffleKey = keyGen.generateKey();  
 52       TokenCache.setShuffleSecretKey(shuffleKey.getEncoded(),  
 53           job.getCredentials());  
 54     }  
 55     //向集群中拷贝所需文件，下面会单独分析（1）  
 56     copyAndConfigureFiles(job, submitJobDir);  
 57     Path submitJobFile = JobSubmissionFiles.getJobConfPath(submitJobDir);  
 58       
 59     // 写分片文件job.split job.splitmetainfo，具体写入过程与MR1相同，可参考以前文章  
 60     LOG.debug("Creating splits at " + jtFs.makeQualified(submitJobDir));  
 61     int maps = writeSplits(job, submitJobDir);  
 62     conf.setInt(MRJobConfig.NUM_MAPS, maps);  
 63     LOG.info("number of splits:" + maps);  
 64   
 65   
 66     // write "queue admins of the queue to which job is being submitted"  
 67     // to job file.  
 68     //设置队列名  
 69     String queue = conf.get(MRJobConfig.QUEUE_NAME,  
 70         JobConf.DEFAULT_QUEUE_NAME);  
 71     AccessControlList acl = submitClient.getQueueAdmins(queue);  
 72     conf.set(toFullPropertyName(queue,  
 73         QueueACL.ADMINISTER_JOBS.getAclName()), acl.getAclString());  
 74   
 75   
 76     // removing jobtoken referrals before copying the jobconf to HDFS  
 77     // as the tasks don't need this setting, actually they may break  
 78     // because of it if present as the referral will point to a  
 79     // different job.  
 80     TokenCache.cleanUpTokenReferral(conf);  
 81   
 82   
 83     if (conf.getBoolean(  
 84         MRJobConfig.JOB_TOKEN_TRACKING_IDS_ENABLED,  
 85         MRJobConfig.DEFAULT_JOB_TOKEN_TRACKING_IDS_ENABLED)) {  
 86       // Add HDFS tracking ids  
 87       ArrayList<String> trackingIds = new ArrayList<String>();  
 88       for (Token<? extends TokenIdentifier> t :  
 89           job.getCredentials().getAllTokens()) {  
 90         trackingIds.add(t.decodeIdentifier().getTrackingId());  
 91       }  
 92       conf.setStrings(MRJobConfig.JOB_TOKEN_TRACKING_IDS,  
 93           trackingIds.toArray(new String[trackingIds.size()]));  
 94     }  
 95   
 96   
 97     // Write job file to submit dir  
 98     //写入job.xml  
 99     writeConf(conf, submitJobFile);  
100       
101     //  
102     // Now, actually submit the job (using the submit name)  
103     //这里才开始真正提交，见下面分析(2)  
104     printTokens(jobId, job.getCredentials());  
105     status = submitClient.submitJob(  
106         jobId, submitJobDir.toString(), job.getCredentials());  
107     if (status != null) {  
108       return status;  
109     } else {  
110       throw new IOException("Could not launch job");  
111     }  
112   } finally {  
113     if (status == null) {  
114       LOG.info("Cleaning up the staging area " + submitJobDir);  
115       if (jtFs != null && submitJobDir != null)  
116         jtFs.delete(submitJobDir, true);  
117   
118   
119     }  
120   }  
121 }

洋洋洒洒一百余行
（这个可谓任务提交的核心部分，前面的都是铺垫）

Step1:

检查job的输出路径是否存在，如果存在则抛出异常。

Step2:

初始化用于存放Job相关资源的路径。

Step3:

设置客户端的host属性： mapreduce.job.submithostname和mapreduce.job.submithostaddress。

Step4:

通过RPC，向Yarn的ResourceManager申请JobID对象。

Step5:

从HDFS的NameNode获取验证用的Token，并将其放入缓存。

Step6:

将作业文件上传到HDFS，这里如果我们前面没有对Job命名的话，默认的名称就会在这里设置成jar的名字。并且，作业默认的副本数是10，如果属性 mapreduce.client.submit.file.replication没有被设置的话。

Step7:

文件上传到HDFS之后，还要被DistributedCache进行缓存起来。这是因为计算节点收到该作业的第一个任务后，就会有DistributedCache自动将作业文件Cache到节点本地目录下，并且会对压缩文件进行解压，如：.zip，.jar，.tar等等，然后开始任务。

最后，对于同一个计算节点接下来收到的任务， DistributedCache不会重复去下载作业文件，而是直接运行任务。如果一个作业的任务数很多，这种设计避免了在同一个节点上对用一个job的文件会下载多次，大大提高了任务运行的效率。

Step8:

对每个输入文件进行split划分。
关于split的分片： http://www.cnblogs.com/admln/p/hadoop-mapper-numbers-question.html

Step9:

将split信息和SplitMetaInfo都写入HDFS中

Step10:

对Map数目设置，上面获得到的split的个数就是实际的Map任务的数目。

Step11:

hadoop2 作业执行过程之作业提交

猜你喜欢