hadoop源码分析：Map-Reduce的过程解析

Posted 2020-07-25 刘超的通俗云计算

一、客户端

Map-Reduce的过程首先是由客户端提交一个任务开始的。

提交任务主要是通过JobClient.runJob(JobConf)静态函数实现的：

public static RunningJob runJob(JobConf job) throws IOException {   //首先生成一个JobClient对象   JobClient jc = new JobClient(job);   ……   //调用submitJob来提交一个任务   running = jc.submitJob(job);   JobID jobId = running.getID();   ……   while (true) {      //while循环中不断得到此任务的状态，并打印到客户端console中   }   return running; }

其中JobClient的submitJob函数实现如下：

public RunningJob submitJob(JobConf job) throws FileNotFoundException,                                 InvalidJobConfException, IOException {   //从JobTracker得到当前任务的id   JobID jobId = jobSubmitClient.getNewJobId();   //准备将任务运行所需要的要素写入HDFS：   //任务运行程序所在的jar封装成job.jar   //任务所要处理的input split信息写入job.split   //任务运行的配置项汇总写入job.xml   Path submitJobDir = new Path(getSystemDir(), jobId.toString());   Path submitJarFile = new Path(submitJobDir, "job.jar");   Path submitSplitFile = new Path(submitJobDir, "job.split");   //此处将-libjars命令行指定的jar上传至HDFS   configureCommandLineOptions(job, submitJobDir, submitJarFile);   Path submitJobFile = new Path(submitJobDir, "job.xml");   ……   //通过input format的格式获得相应的input split，默认类型为FileSplit   InputSplit[] splits =     job.getInputFormat().getSplits(job, job.getNumMapTasks());     // 生成一个写入流，将input split得信息写入job.split文件   FSDataOutputStream out = FileSystem.create(fs,       submitSplitFile, new FsPermission(JOB_FILE_PERMISSION));   try {     //写入job.split文件的信息包括：split文件头，split文件版本号，split的个数，接着依次写入每一个input split的信息。     //对于每一个input split写入：split类型名(默认FileSplit)，split的大小，split的内容(对于FileSplit，写入文件名，此split在文件中的起始位置)，split的location信息(即在那个DataNode上)。     writeSplitsFile(splits, out);   } finally {     out.close();   }   job.set("mapred.job.split.file", submitSplitFile.toString());   //根据split的个数设定map task的个数   job.setNumMapTasks(splits.length);   // 写入job的配置信息入job.xml文件         out = FileSystem.create(fs, submitJobFile,       new FsPermission(JOB_FILE_PERMISSION));   try {     job.writeXml(out);   } finally {     out.close();   }   //真正的调用JobTracker来提交任务   JobStatus status = jobSubmitClient.submitJob(jobId);   …… }

 

二、JobTracker

JobTracker作为一个单独的JVM运行，其运行的main函数主要调用有下面两部分：

• 调用静态函数startTracker(new JobConf())创建一个JobTracker对象
• 调用JobTracker.offerService()函数提供服务

在JobTracker的构造函数中，会生成一个taskScheduler成员变量，来进行Job的调度，默认为JobQueueTaskScheduler，也即按照FIFO的方式调度任务。

在offerService函数中，则调用taskScheduler.start()，在这个函数中，为JobTracker(也即taskScheduler的taskTrackerManager)注册了两个Listener：

• JobQueueJobInProgressListener jobQueueJobInProgressListener用于监控job的运行状态
• EagerTaskInitializationListener eagerTaskInitializationListener用于对Job进行初始化

EagerTaskInitializationListener中有一个线程JobInitThread，不断得到jobInitQueue中的JobInProgress对象，调用JobInProgress对象的initTasks函数对任务进行初始化操作。

在上一节中，客户端调用了JobTracker.submitJob函数，此函数首先生成一个JobInProgress对象，然后调用addJob函数，其中有如下的逻辑：

synchronized (jobs) {   synchronized (taskScheduler) {     jobs.put(job.getProfile().getJobID(), job);     //对JobTracker的每一个listener都调用jobAdded函数     for (JobInProgressListener listener : jobInProgressListeners) {       listener.jobAdded(job);     }   } }

 

EagerTaskInitializationListener的jobAdded函数就是向jobInitQueue中添加一个JobInProgress对象，于是自然触发了此Job的初始化操作，由JobInProgress得initTasks函数完成：

public synchronized void initTasks() throws IOException {   ……   //从HDFS中读取job.split文件从而生成input splits   String jobFile = profile.getJobFile();   Path sysDir = new Path(this.jobtracker.getSystemDir());   FileSystem fs = sysDir.getFileSystem(conf);   DataInputStream splitFile =     fs.open(new Path(conf.get("mapred.job.split.file")));   JobClient.RawSplit[] splits;   try {     splits = JobClient.readSplitFile(splitFile);   } finally {     splitFile.close();   }   //map task的个数就是input split的个数   numMapTasks = splits.length;   //为每个map tasks生成一个TaskInProgress来处理一个input split   maps = new TaskInProgress[numMapTasks];   for(int i=0; i < numMapTasks; ++i) {     inputLength += splits[i].getDataLength();     maps[i] = new TaskInProgress(jobId, jobFile,                                  splits[i],                                  jobtracker, conf, this, i);   }   //对于map task，将其放入nonRunningMapCache，是一个Map<Node, List<TaskInProgress>>，也即对于map task来讲，其将会被分配到其input split所在的Node上。nonRunningMapCache将在JobTracker向TaskTracker分配map task的时候使用。   if (numMapTasks > 0) {      nonRunningMapCache = createCache(splits, maxLevel);   }     //创建reduce task   this.reduces = new TaskInProgress[numReduceTasks];   for (int i = 0; i < numReduceTasks; i++) {     reduces[i] = new TaskInProgress(jobId, jobFile,                                     numMapTasks, i,                                     jobtracker, conf, this);     //reduce task放入nonRunningReduces，其将在JobTracker向TaskTracker分配reduce task的时候使用。     nonRunningReduces.add(reduces[i]);   }     //创建两个cleanup task，一个用来清理map，一个用来清理reduce.   cleanup = new TaskInProgress[2];   cleanup[0] = new TaskInProgress(jobId, jobFile, splits[0],           jobtracker, conf, this, numMapTasks);   cleanup[0].setJobCleanupTask();   cleanup[1] = new TaskInProgress(jobId, jobFile, numMapTasks,                      numReduceTasks, jobtracker, conf, this);   cleanup[1].setJobCleanupTask();   //创建两个初始化 task，一个初始化map，一个初始化reduce.   setup = new TaskInProgress[2];   setup[0] = new TaskInProgress(jobId, jobFile, splits[0],           jobtracker, conf, this, numMapTasks + 1 );   setup[0].setJobSetupTask();   setup[1] = new TaskInProgress(jobId, jobFile, numMapTasks,                      numReduceTasks + 1, jobtracker, conf, this);   setup[1].setJobSetupTask();   tasksInited.set(true);//初始化完毕   …… }

 

三、TaskTracker

TaskTracker也是作为一个单独的JVM来运行的，在其main函数中，主要是调用了new TaskTracker(conf).run()，其中run函数主要调用了：

State offerService() throws Exception {   long lastHeartbeat = 0;   //TaskTracker进行是一直存在的   while (running && !shuttingDown) {       ……       long now = System.currentTimeMillis();       //每隔一段时间就向JobTracker发送heartbeat       long waitTime = heartbeatInterval - (now - lastHeartbeat);       if (waitTime > 0) {         synchronized(finishedCount) {           if (finishedCount[0] == 0) {             finishedCount.wait(waitTime);           }           finishedCount[0] = 0;         }       }       ……       //发送Heartbeat到JobTracker，得到response       HeartbeatResponse heartbeatResponse = transmitHeartBeat(now);       ……      //从Response中得到此TaskTracker需要做的事情       TaskTrackerAction[] actions = heartbeatResponse.getActions();       ……       if (actions != null){         for(TaskTrackerAction action: actions) {           if (action instanceof LaunchTaskAction) {             //如果是运行一个新的Task，则将Action添加到任务队列中             addToTaskQueue((LaunchTaskAction)action);           } else if (action instanceof CommitTaskAction) {             CommitTaskAction commitAction = (CommitTaskAction)action;             if (!commitResponses.contains(commitAction.getTaskID())) {               commitResponses.add(commitAction.getTaskID());             }           } else {             tasksToCleanup.put(action);           }         }       }   }   return State.NORMAL; }

其中transmitHeartBeat主要逻辑如下：

private HeartbeatResponse transmitHeartBeat(long now) throws IOException {   //每隔一段时间，在heartbeat中要返回给JobTracker一些统计信息   boolean sendCounters;   if (now > (previousUpdate + COUNTER_UPDATE_INTERVAL)) {     sendCounters = true;     previousUpdate = now;   }   else {     sendCounters = false;   }   ……   //报告给JobTracker，此TaskTracker的当前状态   if (status == null) {     synchronized (this) {       status = new TaskTrackerStatus(taskTrackerName, localHostname,                                      httpPort,                                      cloneAndResetRunningTaskStatuses(                                        sendCounters),                                      failures,                                      maxCurrentMapTasks,                                      maxCurrentReduceTasks);     }   }   ……   //当满足下面的条件的时候，此TaskTracker请求JobTracker为其分配一个新的Task来运行：   //当前TaskTracker正在运行的map task的个数小于可以运行的map task的最大个数   //当前TaskTracker正在运行的reduce task的个数小于可以运行的reduce task的最大个数   boolean askForNewTask;   long localMinSpaceStart;   synchronized (this) {     askForNewTask = (status.countMapTasks() < maxCurrentMapTasks ||                      status.countReduceTasks() < maxCurrentReduceTasks) &&                     acceptNewTasks;     localMinSpaceStart = minSpaceStart;   }   ……   //向JobTracker发送heartbeat，这是一个RPC调用   HeartbeatResponse heartbeatResponse = jobClient.heartbeat(status,                                                             justStarted, askForNewTask,                                                             heartbeatResponseId);   ……   return heartbeatResponse; }

 

四、JobTracker

当JobTracker被RPC调用来发送heartbeat的时候，JobTracker的heartbeat(TaskTrackerStatus status,boolean initialContact, boolean acceptNewTasks, short responseId)函数被调用：

public synchronized HeartbeatResponse heartbeat(TaskTrackerStatus status,                                                 boolean initialContact, boolean acceptNewTasks, short responseId)   throws IOException {   ……   String trackerName = status.getTrackerName();   ……   short newResponseId = (short)(responseId + 1);   ……   HeartbeatResponse response = new HeartbeatResponse(newResponseId, null);   List<TaskTrackerAction> actions = new ArrayList<TaskTrackerAction>();   //如果TaskTracker向JobTracker请求一个task运行   if (acceptNewTasks) {     TaskTrackerStatus taskTrackerStatus = getTaskTracker(trackerName);     if (taskTrackerStatus == null) {       LOG.warn("Unknown task tracker polling; ignoring: " + trackerName);     } else {       //setup和cleanup的task优先级最高       List<Task> tasks = getSetupAndCleanupTasks(taskTrackerStatus);       if (tasks == null ) {         //任务调度器分配任务         tasks = taskScheduler.assignTasks(taskTrackerStatus);       }       if (tasks != null) {         for (Task task : tasks) {           //将任务放入actions列表，返回给TaskTracker           expireLaunchingTasks.addNewTask(task.getTaskID());           actions.add(new LaunchTaskAction(task));         }       }     }   }   ……   int nextInterval = getNextHeartbeatInterval();   response.setHeartbeatInterval(nextInterval);   response.setActions(                       actions.toArray(new TaskTrackerAction[actions.size()]));   ……   return response; }

默认的任务调度器为JobQueueTaskScheduler，其assignTasks如下：

public synchronized List<Task> assignTasks(TaskTrackerStatus taskTracker)     throws IOException {   ClusterStatus clusterStatus = taskTrackerManager.getClusterStatus();   int numTaskTrackers = clusterStatus.getTaskTrackers();   Collection<JobInProgress> jobQueue =　jobQueueJobInProgressListener.getJobQueue();   int maxCurrentMapTasks = taskTracker.getMaxMapTasks();   int maxCurrentReduceTasks = taskTracker.getMaxReduceTasks();   int numMaps = taskTracker.countMapTasks();   int numReduces = taskTracker.countReduceTasks();   //计算剩余的map和reduce的工作量：remaining   int remainingReduceLoad = 0;   int remainingMapLoad = 0;   synchronized (jobQueue) {     for (JobInProgress job : jobQueue) {       if (job.getStatus().getRunState() == JobStatus.RUNNING) {         int totalMapTasks = job.desiredMaps();         int totalReduceTasks = job.desiredReduces();         remainingMapLoad += (totalMapTasks - job.finishedMaps());         remainingReduceLoad += (totalReduceTasks - job.finishedReduces());       }     }   }   //计算平均每个TaskTracker应有的工作量，remaining/numTaskTrackers是剩余的工作量除以TaskTracker的个数。   int maxMapLoad = 0;   int maxReduceLoad = 0;   if (numTaskTrackers > 0) {     maxMapLoad = Math.min(maxCurrentMapTasks,                           (int) Math.ceil((double) remainingMapLoad /                                           numTaskTrackers));     maxReduceLoad = Math.min(maxCurrentReduceTasks,                              (int) Math.ceil((double) remainingReduceLoad                                              / numTaskTrackers));   }   ……     //map优先于reduce，当TaskTracker上运行的map task数目小于平均的工作量，则向其分配map task   if (numMaps < maxMapLoad) {     int totalNeededMaps = 0;     synchronized (jobQueue) {       for (JobInProgress job : jobQueue) {         if (job.getStatus().getRunState() != JobStatus.RUNNING) {           continue;         }         Task t = job.obtainNewMapTask(taskTracker, numTaskTrackers,             taskTrackerManager.getNumberOfUniqueHosts());         if (t != null) {           return Collections.singletonList(t);         }         ……       }     }   }   //分配完map task，再分配reduce task   if (numReduces < maxReduceLoad) {     int totalNeededReduces = 0;     synchronized (jobQueue) {       for (JobInProgress job : jobQueue) {         if (job.getStatus().getRunState() != JobStatus.RUNNING ||             job.numReduceTasks == 0) {           continue;         }         Task t = job.obtainNewReduceTask(taskTracker, numTaskTrackers,             taskTrackerManager.getNumberOfUniqueHosts());         if (t != null) {           return Collections.singletonList(t);         }         ……       }     }   }   return null; }

从上面的代码中我们可以知道，JobInProgress的obtainNewMapTask是用来分配map task的，其主要调用findNewMapTask，根据TaskTracker所在的Node从nonRunningMapCache中查找TaskInProgress。JobInProgress的obtainNewReduceTask是用来分配reduce task的，其主要调用findNewReduceTask，从nonRunningReduces查找TaskInProgress。

 

五、TaskTracker

在向JobTracker发送heartbeat后，返回的reponse中有分配好的任务LaunchTaskAction，将其加入队列，调用addToTaskQueue，如果是map task则放入mapLancher(类型为TaskLauncher)，如果是reduce task则放入reduceLancher(类型为TaskLauncher)：

private void addToTaskQueue(LaunchTaskAction action) {   if (action.getTask().isMapTask()) {     mapLauncher.addToTaskQueue(action);   } else {     reduceLauncher.addToTaskQueue(action);   } }

TaskLauncher是一个线程，其run函数从上面放入的queue中取出一个TaskInProgress，然后调用startNewTask(TaskInProgress tip)来启动一个task，其又主要调用了localizeJob(TaskInProgress tip)：

private void localizeJob(TaskInProgress tip) throws IOException {   //首先要做的一件事情是有关Task的文件从HDFS拷贝的TaskTracker的本地文件系统中：job.split，job.xml以及job.jar   Path localJarFile = null;   Task t = tip.getTask();   JobID jobId = t.getJobID();   Path jobFile = new Path(t.getJobFile());   ……   Path localJobFile = lDirAlloc.getLocalPathForWrite(                                   getLocalJobDir(jobId.toString())                                   + Path.SEPARATOR + "job.xml",                                   jobFileSize, fConf);   RunningJob rjob = addTaskToJob(jobId, tip);   synchronized (rjob) {     if (!rjob.localized) {       FileSystem localFs = FileSystem.getLocal(fConf);       Path jobDir = localJobFile.getParent();       ……       //将job.split拷贝到本地       systemFS.copyToLocalFile(jobFile, localJobFile);       JobConf localJobConf = new JobConf(localJobFile);       Path workDir = lDirAlloc.getLocalPathForWrite(                        (getLocalJobDir(jobId.toString())                        + Path.SEPARATOR + "work"), fConf);       if (!localFs.mkdirs(workDir)) {         throw new IOException("Mkdirs failed to create "                     + workDir.toString());       }       System.setProperty("job.local.dir", workDir.toString());       localJobConf.set("job.local.dir", workDir.toString());       // copy Jar file to the local FS and unjar it.       String jarFile = localJobConf.getJar();       long jarFileSize = -1;       if (jarFile != null) {         Path jarFilePath = new Path(jarFile);         localJarFile = new Path(lDirAlloc.getLocalPathForWrite(                                    getLocalJobDir(jobId.toString())                                    + Path.SEPARATOR + "jars",                                    5 * jarFileSize, fConf), "job.jar");         if (!localFs.mkdirs(localJarFile.getParent())) {           throw new IOException("Mkdirs failed to create jars directory ");         }         //将job.jar拷贝到本地         systemFS.copyToLocalFile(jarFilePath, localJarFile);         localJobConf.setJar(localJarFile.toString());        //将job得configuration写成job.xml         OutputStream out = localFs.create(localJobFile);         try {           localJobConf.writeXml(out);         } finally {           out.close();         }         // 解压缩job.jar         RunJar.unJar(new File(localJarFile.toString()),                      new File(localJarFile.getParent().toString()));       }       rjob.localized = true;       rjob.jobConf = localJobConf;     }   }   //真正的启动此Task   launchTaskForJob(tip, new JobConf(rjob.jobConf)); }

当所有的task运行所需要的资源都拷贝到本地后，则调用launchTaskForJob，其又调用TaskInProgress的launchTask函数：

public synchronized void launchTask() throws IOException {     ……     //创建task运行目录     localizeTask(task);     if (this.taskStatus.getRunState() == TaskStatus.State.UNASSIGNED) {       this.taskStatus.setRunState(TaskStatus.State.RUNNING);     }     //创建并启动TaskRunner，对于MapTask，创建的是MapTaskRunner，对于ReduceTask，创建的是ReduceTaskRunner     this.runner = task.createRunner(TaskTracker.this, this);     this.runner.start();     this.taskStatus.setStartTime(System.currentTimeMillis()); }

TaskRunner是一个线程，其run函数如下：

public final void run() {     ……     TaskAttemptID taskid = t.getTaskID();     LocalDirAllocator lDirAlloc = new LocalDirAllocator("mapred.local.dir");     File jobCacheDir = null;     if (conf.getJar() != null) {       jobCacheDir = new File(                         new Path(conf.getJar()).getParent().toString());     }     File workDir = new File(lDirAlloc.getLocalPathToRead(                               TaskTracker.getLocalTaskDir(                                 t.getJobID().toString(),                                 t.getTaskID().toString(),                                 t.isTaskCleanupTask())                               + Path.SEPARATOR + MRConstants.WORKDIR,                               conf). toString());     FileSystem fileSystem;     Path localPath;     ……     //拼写classpath     String baseDir;     String sep = System.getProperty("path.separator");     StringBuffer classPath = new StringBuffer();     // start with same classpath as parent process     classPath.append(System.getProperty("java.class.path"));     classPath.append(sep);     if (!workDir.mkdirs()) {       if (!workDir.isDirectory()) {         LOG.fatal("Mkdirs failed to create " + workDir.toString());       }     }     String jar = conf.getJar();     if (jar != null) {             // if jar exists, it into workDir       File[] libs = new File(jobCacheDir, "lib").listFiles();       if (libs != null) {         for (int i = 0; i < libs.length; i++) {           classPath.append(sep);            // add libs from jar to classpath           classPath.append(libs[i]);         }       }       classPath.append(sep);       classPath.append(new File(jobCacheDir, "classes"));       classPath.append(sep);       classPath.append(jobCacheDir);     }     ……     classPath.append(sep);     classPath.append(workDir);     //拼写命令行java及其参数     Vector<String> vargs = new Vector<String>(8);     File jvm =       new File(new File(System.getProperty("java.home"), "bin"), "java");     vargs.add(jvm.toString());     String javaOpts = conf.get("mapred.child.java.opts", "-Xmx200m");     javaOpts = javaOpts.replace("@taskid@", taskid.toString());     String [] javaOptsSplit = javaOpts.split(" ");     String libraryPath = System.getProperty("java.library.path");     if (libraryPath == null) {       libraryPath = workDir.getAbsolutePath();     } else {       libraryPath += sep + workDir;     }     boolean hasUserLDPath = false;     for(int i=0; i<javaOptsSplit.length ;i++) {       if(javaOptsSplit[i].startsWith("-Djava.library.path=")) {         javaOptsSplit[i] += sep + libraryPath;         hasUserLDPath = true;         break;       }     }     if(!hasUserLDPath) {       vargs.add("-Djava.library.path=" + libraryPath);     }     for (int i = 0; i < javaOptsSplit.length; i++) {       vargs.add(javaOptsSplit[i]);     }     //添加Child进程的临时文件夹     String tmp = conf.get("mapred.child.tmp", "./tmp");     Path tmpDir = new Path(tmp);     if (!tmpDir.isAbsolute()) {       tmpDir = new Path(workDir.toString(), tmp);     }     FileSystem localFs = FileSystem.getLocal(conf);     if (!localFs.mkdirs(tmpDir) && !localFs.getFileStatus(tmpDir).isDir()) {       throw new IOException("Mkdirs failed to create " + tmpDir.toString());     }     vargs.add("-Djava.io.tmpdir=" + tmpDir.toString());     // Add classpath.     vargs.add("-classpath");     vargs.add(classPath.toString());     //log文件夹     long logSize = TaskLog.getTaskLogLength(conf);     vargs.add("-Dhadoop.log.dir=" +         new File(System.getProperty("hadoop.log.dir")         ).getAbsolutePath());     vargs.add("-Dhadoop.root.logger=INFO,TLA");     vargs.add("-Dhadoop.tasklog.taskid=" + taskid);     vargs.add("-Dhadoop.tasklog.totalLogFileSize=" + logSize);     // 运行map task和reduce task的子进程的main class是Child     vargs.add(Child.class.getName());  // main of Child     ……     //运行子进程     jvmManager.launchJvm(this,         jvmManager.constructJvmEnv(setup,vargs,stdout,stderr,logSize,             workDir, env, pidFile, conf)); }

 

六、Child

真正的map task和reduce task都是在Child进程中运行的，Child的main函数的主要逻辑如下：

while (true) {   //从TaskTracker通过网络通信得到JvmTask对象   JvmTask myTask = umbilical.getTask(jvmId);   ……   idleLoopCount = 0;   task = myTask.getTask();   taskid = task.getTaskID();   isCleanup = task.isTaskCleanupTask();   JobConf job = new JobConf(task.getJobFile());   TaskRunner.setupWorkDir(job);   numTasksToExecute = job.getNumTasksToExecutePerJvm();   task.setConf(job);   defaultConf.addResource(new Path(task.getJobFile()));   ……   //运行task   task.run(job, umbilical);             // run the task   if (numTasksToExecute > 0 && ++numTasksExecuted == numTasksToExecute) {     break;   } }

6.1、MapTask

如果task是MapTask，则其run函数如下：

public void run(final JobConf job, final TaskUmbilicalProtocol umbilical)   throws IOException {   //用于同TaskTracker进行通信，汇报运行状况   final Reporter reporter = getReporter(umbilical);   startCommunicationThread(umbilical);   initialize(job, reporter);   ……   //map task的输出   int numReduceTasks = conf.getNumReduceTasks();   MapOutputCollector collector = null;   if (numReduceTasks > 0) {     collector = new MapOutputBuffer(umbilical, job, reporter);   } else {     collector = new DirectMapOutputCollector(umbilical, job, reporter);   }   //读取input split，按照其中的信息，生成RecordReader来读取数据 instantiatedSplit = (InputSplit)       ReflectionUtils.newInstance(job.getClassByName(splitClass), job);   DataInputBuffer splitBuffer = new DataInputBuffer();   splitBuffer.reset(split.getBytes(), 0, split.getLength());   instantiatedSplit.readFields(splitBuffer);   if (instantiatedSplit instanceof FileSplit) {     FileSplit fileSplit = (FileSplit) instantiatedSplit;     job.set("map.input.file", fileSplit.getPath().toString());     job.setLong("map.input.start", fileSplit.getStart());     job.setLong("map.input.length", fileSplit.getLength());   }   RecordReader rawIn =                  // open input     job.getInputFormat().getRecordReader(instantiatedSplit, job, reporter);   RecordReader in = isSkipping() ?       new SkippingRecordReader(rawIn, getCounters(), umbilical) :       new TrackedRecordReader(rawIn, getCounters());   job.setBoolean("mapred.skip.on", isSkipping());   //对于map task，生成一个MapRunnable，默认是MapRunner   MapRunnable runner =     ReflectionUtils.newInstance(job.getMapRunnerClass(), job);   try {     //MapRunner的run函数就是依次读取RecordReader中的数据，然后调用Mapper的map函数进行处理。     runner.run(in, collector, reporter);          collector.flush();   } finally {     in.close();                               // close input     collector.close();   }   done(umbilical); }

MapRunner的run函数就是依次读取RecordReader中的数据，然后调用Mapper的map函数进行处理：

public void run(RecordReader<K1, V1> input, OutputCollector<K2, V2> output,                 Reporter reporter)   throws IOException {   try {     K1 key = input.createKey();     V1 value = input.createValue();     while (input.next(key, value)) {       mapper.map(key, value, output, reporter);       if(incrProcCount) {         reporter.incrCounter(SkipBadRecords.COUNTER_GROUP,             SkipBadRecords.COUNTER_MAP_PROCESSED_RECORDS, 1);       }     }   } finally {     mapper.close();   } }

结果集全部收集到MapOutputBuffer中，其collect函数如下：

public synchronized void collect(K key, V value)     throws IOException {   reporter.progress();   ……   //从此处看，此buffer是一个ring的数据结构   final int kvnext = (kvindex + 1) % kvoffsets.length;   spillLock.lock();   try {     boolean kvfull;     do {       //在ring中，如果下一个空闲位置接上起始位置的话，则表示满了       kvfull = kvnext == kvstart;       //在ring中计算是否需要将buffer写入硬盘的阈值       final boolean kvsoftlimit = ((kvnext > kvend)           ? kvnext - kvend > softRecordLimit           : kvend - kvnext <= kvoffsets.length - softRecordLimit);       //如果到达阈值，则开始将buffer写入硬盘，写成spill文件。       //startSpill主要是notify一个背后线程SpillThread的run()函数，开始调用sortAndSpill()开始排序，合并，写入硬盘       if (kvstart == kvend && kvsoftlimit) {         startSpill();       }       //如果buffer满了，则只能等待写入完毕       if (kvfull) {           while (kvstart != kvend) {             reporter.progress();             spillDone.await();           }       }     } while (kvfull);   } finally {     spillLock.unlock();   }   try {     //如果buffer不满，则将key, value写入buffer     int keystart = bufindex;     keySerializer.serialize(key);     final int valstart = bufindex;     valSerializer.serialize(value);     int valend = bb.markRecord();     //调用设定的partitioner，根据key, value取得partition id     final int partition = partitioner.getPartition(key, value, partitions);     mapOutputRecordCounter.increment(1);     mapOutputByteCounter.increment(valend >= keystart         ? valend - keystart         : (bufvoid - keystart) + valend);     //将parition id以及key, value在buffer中的偏移量写入索引数组     int ind = kvindex * ACCTSIZE;     kvoffsets[kvindex] = ind;     kvindices[ind + PARTITION] = partition;     kvindices[ind + KEYSTART] = keystart;     kvindices[ind + VALSTART] = valstart;     kvindex = kvnext;   } catch (MapBufferTooSmallException e) {     LOG.info("Record too large for in-memory buffer: " + e.getMessage());     spillSingleRecord(key, value);     mapOutputRecordCounter.increment(1);     return;   } }

内存buffer的格式如下：

(见几位hadoop大侠的分析http://blog.csdn.net/HEYUTAO007/archive/2010/07/10/5725379.aspx 以及http://caibinbupt.javaeye.com/)

kvoffsets是为了写入内存前排序使用的。

从上面可知，内存buffer写入硬盘spill文件的函数为sortAndSpill：

private void sortAndSpill() throws IOException {   ……   FSDataOutputStream out = null;   FSDataOutputStream indexOut = null;   IFileOutputStream indexChecksumOut = null;   //创建硬盘上的spill文件   Path filename = mapOutputFile.getSpillFileForWrite(getTaskID(),                                   numSpills, size);   out = rfs.create(filename);   ……   final int endPosition = (kvend > kvstart)     ? kvend     : kvoffsets.length + kvend;   //按照partition的顺序对buffer中的数据进行排序   sorter.sort(MapOutputBuffer.this, kvstart, endPosition, reporter);   int spindex = kvstart;   InMemValBytes value = new InMemValBytes();   //依次一个一个parition的写入文件   for (int i = 0; i < partitions; ++i) {     IFile.Writer<K, V> writer = null;     long segmentStart = out.getPos();     writer = new Writer<K, V>(job, out, keyClass, valClass, codec);     //如果combiner为空，则直接写入文件     if (null == combinerClass) {         ……         writer.append(key, value);         ++spindex;      }      else {         ……         //如果combiner不为空，则先combine，调用combiner.reduce(…)函数后再写入文件         combineAndSpill(kvIter, combineInputCounter);      }   }   …… }

当map阶段结束的时候，MapOutputBuffer的flush函数会被调用，其也会调用sortAndSpill将buffer中的写入文件，然后再调用mergeParts来合并写入在硬盘上的多个spill:

private void mergeParts() throws IOException {     ……     //对于每一个partition     for (int parts = 0; parts < partitions; parts++){       //create the segments to be merged       List<Segment<K, V>> segmentList =         new ArrayList<Segment<K, V>>(numSpills);       TaskAttemptID mapId = getTaskID();        //依次从各个spill文件中收集属于当前partition的段       for(int i = 0; i < numSpills; i++) {         final IndexRecord indexRecord =           getIndexInformation(mapId, i, parts);         long segmentOffset = indexRecord.startOffset;         long segmentLength = indexRecord.partLength;         Segment<K, V> s =           new Segment<K, V>(job, rfs, filename[i], segmentOffset,                             segmentLength, codec, true);         segmentList.add(i, s);       }       //将属于同一个partition的段merge到一起       RawKeyValueIterator kvIter =         Merger.merge(job, rfs,                      keyClass, valClass,                      segmentList, job.getInt("io.sort.factor", 100),                      new Path(getTaskID().toString()),                      job.getOutputKeyComparator(), reporter);       //写入合并后的段到文件       long segmentStart = finalOut.getPos();       Writer<K, V> writer =           new Writer<K, V>(job, finalOut, keyClass, valClass, codec);       if (null == combinerClass || numSpills < minSpillsForCombine) {         Merger.writeFile(kvIter, writer, reporter, job);       } else {         combineCollector.setWriter(writer);         combineAndSpill(kvIter, combineInputCounter);       }       ……     } }

6.2、ReduceTask

ReduceTask的run函数如下：

public void run(JobConf job, final TaskUmbilicalProtocol umbilical)   throws IOException {   job.setBoolean("mapred.skip.on", isSkipping());   //对于reduce，则包含三个步骤：拷贝，排序，Reduce   if (isMapOrReduce()) {     copyPhase = getProgress().addPhase("copy");     sortPhase  = getProgress().addPhase("sort");     reducePhase = getProgress().addPhase("reduce");   }   startCommunicationThread(umbilical);   final Reporter reporter = getReporter(umbilical);   initialize(job, reporter);   //copy阶段，主要使用ReduceCopier的fetchOutputs函数获得map的输出。创建多个线程MapOutputCopier，其中copyOutput进行拷贝。   boolean isLocal = "local".equals(job.get("mapred.job.tracker", "local"));   if (!isLocal) {     reduceCopier = new ReduceCopier(umbilical, job);     if (!reduceCopier.fetchOutputs()) {         ……     }   }   copyPhase.complete();   //sort阶段，将得到的map输出合并，直到文件数小于io.sort.factor时停止，返回一个Iterator用于访问key-value   setPhase(TaskStatus.Phase.SORT);   statusUpdate(umbilical);   final FileSystem rfs = FileSystem.getLocal(job).getRaw();   RawKeyValueIterator rIter = isLocal     ? Merger.merge(job, rfs, job.getMapOutputKeyClass(),         job.getMapOutputValueClass(), codec, getMapFiles(rfs, true),         !conf.getKeepFailedTaskFiles(), job.getInt("io.sort.factor", 100),         new Path(getTaskID().toString()), job.getOutputKeyComparator(),         reporter)     : reduceCopier.createKVIterator(job, rfs, reporter);   mapOutputFilesOnDisk.clear();   sortPhase.complete();   //reduce阶段   setPhase(TaskStatus.Phase.REDUCE);   ……   Reducer reducer = ReflectionUtils.newInstance(job.getReducerClass(), job);   Class keyClass = job.getMapOutputKeyClass();   Class valClass = job.getMapOutputValueClass();   ReduceValuesIterator values = isSkipping() ?      new SkippingReduceValuesIterator(rIter,           job.getOutputValueGroupingComparator(), keyClass, valClass,           job, reporter, umbilical) :       new ReduceValuesIterator(rIter,       job.getOutputValueGroupingComparator(), keyClass, valClass,       job, reporter);   //逐个读出key-value list，然后调用Reducer的reduce函数   while (values.more()) {     reduceInputKeyCounter.increment(1);     reducer.reduce(values.getKey(), values, collector, reporter);     values.nextKey();     values.informReduceProgress();   }   reducer.close();   out.close(reporter);   done(umbilical); }

 

七、总结

Map-Reduce的过程总结如下图：