博客
关于我
强烈建议你试试无所不能的chatGPT,快点击我
监听器初始化Job、JobTracker相应TaskTracker心跳、调度器分配task源码级分析
阅读量:5235 次
发布时间:2019-06-14

本文共 31929 字,大约阅读时间需要 106 分钟。

  JobTracker和TaskTracker分别启动之后(,),taskTracker会通过心跳与JobTracker通信,并获取分配它的任务。用户将作业提交到JobTracker之后,放入相应的数据结构中,静等被分配。这篇文章已经分析了用户提交作业的最后步骤,主要是构造作业对应的JobInProgress并加入jobs,告知所有的JobInProgressListener。

  默认调度器创建了两个Listener:JobQueueJobInProgressListener和EagerTaskInitializationListener,用户提交的作业被封装成JobInProgress job加入这两个Listener。

  一、JobQueueJobInProgressListener.jobAdded(job)会将此JobInProgress放入Map<JobSchedulingInfo, JobInProgress> jobQueue中。 

  二、EagerTaskInitializationListener.jobAdded(job)会将此JobInProgress放入List<JobInProgress> jobInitQueue中,然后调用resortInitQueue()对这个列表进行排序先按优先级相同则按开始时间;然后唤醒在此对象监视器上等待的所有线程jobInitQueue.notifyAll()。EagerTaskInitializationListener.start()方法已经在调度器start时运行,会创建一个线程JobInitManager implements Runnable,它的run方法主要是监控jobInitQueue列表,一旦发现不为空就获取第一个JobInProgress,然后创建一个InitJob implements Runnable初始化线程并放入线程池ExecutorService threadPool(这个线程池在构建EagerTaskInitializationListener对象时由构造方法实现),InitJob线程的run方法就一句话ttm.initJob(job),调用的是JobTracker的initJob(job)方法对JIP进行初始化,实际调用JobInProgress.initTasks()对job进行初始化,initTasks()方法代码如下:

1  /**  2    * Construct the splits, etc.  This is invoked from an async  3    * thread so that split-computation doesn't block anyone.  4    */  5   //任务Task分两种: MapTask 和reduceTask,它们的管理对象都是TaskInProgress 。  6   public synchronized void initTasks()   7   throws IOException, KillInterruptedException, UnknownHostException {  8     if (tasksInited || isComplete()) {  9       return; 10     } 11     synchronized(jobInitKillStatus){ 12       if(jobInitKillStatus.killed || jobInitKillStatus.initStarted) { 13         return; 14       } 15       jobInitKillStatus.initStarted = true; 16     } 17  18     LOG.info("Initializing " + jobId); 19     final long startTimeFinal = this.startTime; 20     // log job info as the user running the job 21     try { 22     userUGI.doAs(new PrivilegedExceptionAction() { 23       @Override 24       public Object run() throws Exception { 25         JobHistory.JobInfo.logSubmitted(getJobID(), conf, jobFile,  26             startTimeFinal, hasRestarted()); 27         return null; 28       } 29     }); 30     } catch(InterruptedException ie) { 31       throw new IOException(ie); 32     } 33      34     // log the job priority 35     setPriority(this.priority); 36      37     // 38     // generate security keys needed by Tasks 39     // 40     generateAndStoreTokens(); 41      42     // 43     // read input splits and create a map per a split 44     // 45     TaskSplitMetaInfo[] splits = createSplits(jobId); 46     if (numMapTasks != splits.length) { 47       throw new IOException("Number of maps in JobConf doesn't match number of " + 48               "recieved splits for job " + jobId + "! " + 49               "numMapTasks=" + numMapTasks + ", #splits=" + splits.length); 50     } 51     numMapTasks = splits.length;//map task的个数就是input split的个数 52  53     // Sanity check the locations so we don't create/initialize unnecessary tasks 54     for (TaskSplitMetaInfo split : splits) { 55       NetUtils.verifyHostnames(split.getLocations()); 56     } 57      58     jobtracker.getInstrumentation().addWaitingMaps(getJobID(), numMapTasks); 59     jobtracker.getInstrumentation().addWaitingReduces(getJobID(), numReduceTasks); 60     this.queueMetrics.addWaitingMaps(getJobID(), numMapTasks); 61     this.queueMetrics.addWaitingReduces(getJobID(), numReduceTasks); 62  63     maps = new TaskInProgress[numMapTasks]; //为每个map tasks生成一个TaskInProgress来处理一个input split  64     for(int i=0; i < numMapTasks; ++i) { 65       inputLength += splits[i].getInputDataLength(); 66       maps[i] = new TaskInProgress(jobId, jobFile,         //类型是map task 67                                    splits[i],  68                                    jobtracker, conf, this, i, numSlotsPerMap); 69     } 70     LOG.info("Input size for job " + jobId + " = " + inputLength 71         + ". Number of splits = " + splits.length); 72  73     // Set localityWaitFactor before creating cache 74     localityWaitFactor =  75       conf.getFloat(LOCALITY_WAIT_FACTOR, DEFAULT_LOCALITY_WAIT_FACTOR); 76     /* 对于map task,将其放入nonRunningMapCache,是一个Map
>,也即对于map task来讲,其将会被分配到其input 77 split所在的Node上。在此,Node代表一个datanode或者机架或者数据中 心。nonRunningMapCache将在JobTracker向TaskTracker分配map task的 时候使用。 78 */ 79 if (numMapTasks > 0) { 80 //通过createCache()方法为这些TaskInProgress对象产生一个未执行任务的Map缓存nonRunningMapCache。 81 //slave端的TaskTracker向master发送心跳时,就可以直接从这个cache中取任务去执行。 82 nonRunningMapCache = createCache(splits, maxLevel); 83 } 84 85 // set the launch time 86 this.launchTime = jobtracker.getClock().getTime(); 87 88 // 89 // Create reduce tasks 90 // 91 //其次JobInProgress会创建Reduce的监控对象,这个比较简单,根据JobConf里指定的Reduce数目创建, 92 //缺省只创建1个Reduce任务。监控和调度Reduce任务的是TaskInProgress类,不过构造方法有所不同, 93 //TaskInProgress会根据不同参数分别创建具体的MapTask或者ReduceTask。同样地, 94 //initTasks()也会通过createCache()方法产生nonRunningReduceCache成员。 95 this.reduces = new TaskInProgress[numReduceTasks]; 96 for (int i = 0; i < numReduceTasks; i++) { 97 reduces[i] = new TaskInProgress(jobId, jobFile, //这是reduce task 98 numMapTasks, i, 99 jobtracker, conf, this, numSlotsPerReduce);100 /*reducetask放入nonRunningReduces,其将在JobTracker向TaskTracker分配reduce task的时候使用。*/ 101 nonRunningReduces.add(reduces[i]);102 }103 104 // Calculate the minimum number of maps to be complete before 105 // we should start scheduling reduces106 completedMapsForReduceSlowstart = 107 (int)Math.ceil(108 (conf.getFloat("mapred.reduce.slowstart.completed.maps", 109 DEFAULT_COMPLETED_MAPS_PERCENT_FOR_REDUCE_SLOWSTART) * 110 numMapTasks));111 112 // ... use the same for estimating the total output of all maps113 resourceEstimator.setThreshhold(completedMapsForReduceSlowstart);114 115 // create cleanup two cleanup tips, one map and one reduce.116 //创建两个cleanup task,一个用来清理map,一个用来清理reduce. 117 cleanup = new TaskInProgress[2];118 119 // cleanup map tip. This map doesn't use any splits. Just assign an empty120 // split.121 TaskSplitMetaInfo emptySplit = JobSplit.EMPTY_TASK_SPLIT;122 cleanup[0] = new TaskInProgress(jobId, jobFile, emptySplit, 123 jobtracker, conf, this, numMapTasks, 1);124 cleanup[0].setJobCleanupTask();125 126 // cleanup reduce tip.127 cleanup[1] = new TaskInProgress(jobId, jobFile, numMapTasks,128 numReduceTasks, jobtracker, conf, this, 1);129 cleanup[1].setJobCleanupTask();130 131 // create two setup tips, one map and one reduce.132 //创建两个初始化 task,一个初始化map,一个初始化reduce. 133 setup = new TaskInProgress[2];134 135 // setup map tip. This map doesn't use any split. Just assign an empty136 // split.137 setup[0] = new TaskInProgress(jobId, jobFile, emptySplit, 138 jobtracker, conf, this, numMapTasks + 1, 1);139 setup[0].setJobSetupTask();140 141 // setup reduce tip.142 setup[1] = new TaskInProgress(jobId, jobFile, numMapTasks,143 numReduceTasks + 1, jobtracker, conf, this, 1);144 setup[1].setJobSetupTask();145 146 synchronized(jobInitKillStatus){147 jobInitKillStatus.initDone = true;148 if(jobInitKillStatus.killed) {149 throw new KillInterruptedException("Job " + jobId + " killed in init");150 }151 }152 //JobInProgress创建完TaskInProgress后,最后构造JobStatus并记录job正在执行中,153 //然后再调用JobHistory.JobInfo.logInited()记录job的执行日志。154 tasksInited = true;155 JobHistory.JobInfo.logInited(profile.getJobID(), this.launchTime, 156 numMapTasks, numReduceTasks);157 158 // Log the number of map and reduce tasks159 LOG.info("Job " + jobId + " initialized successfully with " + numMapTasks160 + " map tasks and " + numReduceTasks + " reduce tasks.");161 }
View Code

   initTasks方法的主要工作是读取上传的分片信息,检查分片的有效性及要和配置文件中的numMapTasks相等,然后创建numMapTasks个TaskInProgress作为Map Task。通过createCache方法,将没有找到对应分片的map放入nonLocalMaps中,获取分片所在节点,然后将节点与其上分片对应的map对应起来,放入Map<Node, List<TaskInProgress>> cache之中,需要注意的是还会根据设定的网络深度存储父节点(可能存在多个子节点)下所有子节点包含的map,从这可以看出这里实现了本地化,将这个cache赋值给nonRunningMapCache表示还未运行的map。然后是创建reduce task,创建numReduceTasks个TaskInProgress,放入nonRunningReduces。这里需要注意:map和reduce都是TaskInProgress那以后咋区分呢?其实这两种的构造函数是不同的,判断两种类型的task的根据就是splitInfo有无设置,map task对splitInfo进行了设置,而reduce task则设splitInfo=null。然后是获取map task完成的最小数量才可以调度reduce task。创建两个清理task:cleanup = new TaskInProgress[2],一个用来清理map task(这个也是一个map task),一个用来清理reduce task(这个也是一个reduce task),TaskInProgress构造函数的task个数参数都为1,map的splitInfo是JobSplit.EMPTY_TASK_SPLIT;创建两个初始化task:setup = new TaskInProgress[2],一个用来初始化map task(这个也是一个map task),一个用来初始化reduce task(这个也是一个reduce task),这4个TaskInProgress都会设置对应的标记为来表示类型。最后是设置一个标记位来表示完成初始化工作。

  这样EagerTaskInitializationListener在JobTracker端就完成了对Job的初始化工作,所有task等待taskTracker的心跳被调度。

  来看TaskTracker通过心跳提交状态的方法JobTracker.heartbeat,该方法代码:

1   /**  2    * The periodic heartbeat mechanism between the {
@link TaskTracker} and 3 * the {
@link JobTracker}. 4 * 5 * The {
@link JobTracker} processes the status information sent by the 6 * {
@link TaskTracker} and responds with instructions to start/stop 7 * tasks or jobs, and also 'reset' instructions during contingencies. 8 */ 9 public synchronized HeartbeatResponse heartbeat(TaskTrackerStatus status, 10 boolean restarted, 11 boolean initialContact, 12 boolean acceptNewTasks, 13 short responseId) 14 throws IOException { 15 if (LOG.isDebugEnabled()) { 16 LOG.debug("Got heartbeat from: " + status.getTrackerName() + 17 " (restarted: " + restarted + 18 " initialContact: " + initialContact + 19 " acceptNewTasks: " + acceptNewTasks + ")" + 20 " with responseId: " + responseId); 21 } 22 23 // Make sure heartbeat is from a tasktracker allowed by the jobtracker. 24 if (!acceptTaskTracker(status)) { 25 throw new DisallowedTaskTrackerException(status); 26 } 27 28 // First check if the last heartbeat response got through 29 String trackerName = status.getTrackerName(); 30 long now = clock.getTime(); 31 if (restarted) { 32 faultyTrackers.markTrackerHealthy(status.getHost()); 33 } else { 34 faultyTrackers.checkTrackerFaultTimeout(status.getHost(), now); 35 } 36 37 HeartbeatResponse prevHeartbeatResponse = 38 trackerToHeartbeatResponseMap.get(trackerName); 39 boolean addRestartInfo = false; 40 41 if (initialContact != true) { 42 // If this isn't the 'initial contact' from the tasktracker, 43 // there is something seriously wrong if the JobTracker has 44 // no record of the 'previous heartbeat'; if so, ask the 45 // tasktracker to re-initialize itself. 46 if (prevHeartbeatResponse == null) { 47 // This is the first heartbeat from the old tracker to the newly 48 // started JobTracker 49 if (hasRestarted()) { 50 addRestartInfo = true; 51 // inform the recovery manager about this tracker joining back 52 recoveryManager.unMarkTracker(trackerName); 53 } else { 54 // Jobtracker might have restarted but no recovery is needed 55 // otherwise this code should not be reached 56 LOG.warn("Serious problem, cannot find record of 'previous' " + 57 "heartbeat for '" + trackerName + 58 "'; reinitializing the tasktracker"); 59 return new HeartbeatResponse(responseId, 60 new TaskTrackerAction[] {
new ReinitTrackerAction()}); 61 } 62 63 } else { 64 65 // It is completely safe to not process a 'duplicate' heartbeat from a 66 // {@link TaskTracker} since it resends the heartbeat when rpcs are 67 // lost see {@link TaskTracker.transmitHeartbeat()}; 68 // acknowledge it by re-sending the previous response to let the 69 // {@link TaskTracker} go forward. 70 if (prevHeartbeatResponse.getResponseId() != responseId) { 71 LOG.info("Ignoring 'duplicate' heartbeat from '" + 72 trackerName + "'; resending the previous 'lost' response"); 73 return prevHeartbeatResponse; 74 } 75 } 76 } 77 78 // Process this heartbeat 79 short newResponseId = (short)(responseId + 1);  //响应编号+1 80 status.setLastSeen(now); 81 if (!processHeartbeat(status, initialContact, now)) { 82 if (prevHeartbeatResponse != null) { 83 trackerToHeartbeatResponseMap.remove(trackerName); 84 } 85 return new HeartbeatResponse(newResponseId, 86 new TaskTrackerAction[] {
new ReinitTrackerAction()}); 87 } 88 89 // Initialize the response to be sent for the heartbeat 90 HeartbeatResponse response = new HeartbeatResponse(newResponseId, null); 91 List
actions = new ArrayList
(); 92 boolean isBlacklisted = faultyTrackers.isBlacklisted(status.getHost()); 93 // Check for new tasks to be executed on the tasktracker 94 if (recoveryManager.shouldSchedule() && acceptNewTasks && !isBlacklisted) { 95 TaskTrackerStatus taskTrackerStatus = getTaskTrackerStatus(trackerName); 96 if (taskTrackerStatus == null) { 97 LOG.warn("Unknown task tracker polling; ignoring: " + trackerName); 98 } else { 99 //setup和cleanup的task优先级最高 100 List
tasks = getSetupAndCleanupTasks(taskTrackerStatus);101 if (tasks == null ) {102 //任务调度器分配任务 103 tasks = taskScheduler.assignTasks(taskTrackers.get(trackerName)); //分配任务Map OR Reduce Task104 }105 106 if (tasks != null) {107 for (Task task : tasks) {108 //将任务放入actions列表,返回给TaskTracker109 expireLaunchingTasks.addNewTask(task.getTaskID());110 if(LOG.isDebugEnabled()) {111 LOG.debug(trackerName + " -> LaunchTask: " + task.getTaskID());112 }113 actions.add(new LaunchTaskAction(task));114 }115 }116 }117 }118 119 // Check for tasks to be killed120 List
killTasksList = getTasksToKill(trackerName);121 if (killTasksList != null) {122 actions.addAll(killTasksList);123 }124 125 // Check for jobs to be killed/cleanedup126 List
killJobsList = getJobsForCleanup(trackerName);127 if (killJobsList != null) {128 actions.addAll(killJobsList);129 }130 131 // Check for tasks whose outputs can be saved132 List
commitTasksList = getTasksToSave(status);133 if (commitTasksList != null) {134 actions.addAll(commitTasksList);135 }136 137 // calculate next heartbeat interval and put in heartbeat response138 int nextInterval = getNextHeartbeatInterval();139 response.setHeartbeatInterval(nextInterval);140 response.setActions(141 actions.toArray(new TaskTrackerAction[actions.size()]));142 143 // check if the restart info is req144 if (addRestartInfo) {145 response.setRecoveredJobs(recoveryManager.getJobsToRecover());146 }147 148 // Update the trackerToHeartbeatResponseMap149 trackerToHeartbeatResponseMap.put(trackerName, response);150 151 // Done processing the hearbeat, now remove 'marked' tasks152 removeMarkedTasks(trackerName);153 154 return response;155 }

  一、该方法包括5个参数:A、status封装了TaskTracker上的各种状态信息,包括: TaskTracker名称;TaskTracker主机名;TaskTracker对外的HTTp端口号;该TaskTracker上已经失败的任务总数;正在运行的各个任务的运行状态;上次汇报心跳的时间;Map slot总数,即同时运行的Map Task总数;Reduce slot总数;TaskTracker健康状态;TaskTracker资源(内存、CPU)信息。B、restarted表示TaskTracker是否刚刚重启。C、initialContact表示TaskTracker是否初次链接JobTracker。D、acceptNewTasks表示TaskTracker是否可以接受新的任务,这通常取决于solt是否有剩余和节点的健康状况等。E、responseID表示心跳相应编号,用于防止重复发送心跳,没接收一次心跳后该值加1。

  二、acceptTaskTracker(status)检查心跳是否来自于JobTracker所允许的TaskTracker,当一个TaskTracker在mapred.hosts(include list是合法的节点列表,只有位于该列表中的节点才可以允许JobTracker发起链接请求)指定的主机列表中,不在mapred.exclude(exclude list是一个非法节点列表,所有位于这个列表中的节点将无法与JobTracker链接)指定的主机列表中时,可以接入JobTracker。默认情况下这两个列表都为空,可在配置文件mapred-site.xml中配置,可动态加载。

  三、如果TaskTracker重启了,则将它标注为健康的TaskTracker,并从黑名单(Hadoop允许用户编写一个脚本监控TaskTracker是否健康,并通过心跳将检测结果发送给JobTracker,一旦发现不健康,JobTracker会将该TaskTracker加入黑名单,不再分配任务,直到检测结果为健康)或灰名单(JobTracker会记录每个TaskTracker被作业加入黑名单的次数#backlist,满足一定的要求就加入JobTracker的灰名单)中清除,否则,启动TaskTracker容错机制以检查它是否处于健康状态。

  四、获取该TaskTracker对应的HeartbeatResponse,并检查。如果不是第一次连接JobTracker,且对应的HeartbeatResponse等于null(表明JobTracker没有对应的记录,可能TaskTracker出错也可能JobTracker重启了),如果JobTracker重启了,则从recoveryManager中删除这个trackerName,否则向TaskTracker发送初始化命令ReinitTrackerAction;HeartbeatResponse不等于null,有可能是TaskTracker重复发送心跳,如果是重复发送心跳则返回当前的HeartbeatResponse。

  五、更新响应编号(+1);记录心跳发送时间status.setLastSeen(now);然后调用processHeartbeat(status, initialContact, now)方法来处理TaskTracker发送过来的心跳,先通过updateTaskTrackerStatus方法更新一些资源统计情况,并替换掉旧的taskTracker的状态,如果是初次链接JobTracker且JobTracker中有此taskTracker的记录(TT重启),则需要清空和这个TaskTracker相关的信息,如果不是初次链接JobTracker且JobTracker并没有发现此TaskTracker以前的记录,则直接返回false;如果初次链接JobTracker且包含在黑名单中,则increment the count of blacklisted trackers,然后加入trackerExpiryQueue和hostnameToTaskTracker;updateTaskStatuses(trackerStatus)更新task的状态,这个好复杂留待以后分析;updateNodeHealthStatus(trackerStatus, timeStamp)更新节点健康状态;返回true。若返回false,需要从trackerToHeartbeatResponseMap中删除对应的trackerName信息并返回给TaskTracker初始化命令ReinitTrackerAction。

  六、构造TaskTracker的心跳应答。首先获取setup和cleanup的tasks,如果tasks==null则用调度器(默认是JobQueueTaskScheduler)去分配task,tasks = taskScheduler.assignTasks(taskTrackers.get(trackerName)),会获得Map Task或者Reduce Task,对应assignTasks方法的代码如下:

1 //JobQueueTaskScheduler最重要的方法是assignTasks,他实现了工作调度。  2   @Override  3   public synchronized List
assignTasks(TaskTracker taskTracker) 4 throws IOException { 5 TaskTrackerStatus taskTrackerStatus = taskTracker.getStatus(); 6 ClusterStatus clusterStatus = taskTrackerManager.getClusterStatus(); 7 final int numTaskTrackers = clusterStatus.getTaskTrackers(); 8 final int clusterMapCapacity = clusterStatus.getMaxMapTasks(); 9 final int clusterReduceCapacity = clusterStatus.getMaxReduceTasks(); 10 11 Collection
jobQueue = 12 jobQueueJobInProgressListener.getJobQueue(); 13 //首先它会检查 TaskTracker 端还可以做多少个 map 和 reduce 任务,将要派发的任务数是否超出这个数, 14 //是否超出集群的任务平均剩余可负载数。如果都没超出,则为此TaskTracker 分配一个 MapTask 或 ReduceTask 。 15 // 16 // Get map + reduce counts for the current tracker. 17 // 18 final int trackerMapCapacity = taskTrackerStatus.getMaxMapSlots(); 19 final int trackerReduceCapacity = taskTrackerStatus.getMaxReduceSlots(); 20 final int trackerRunningMaps = taskTrackerStatus.countMapTasks(); 21 final int trackerRunningReduces = taskTrackerStatus.countReduceTasks(); 22 23 // Assigned tasks 24 List
assignedTasks = new ArrayList
(); 25 26 // 27 // Compute (running + pending) map and reduce task numbers across pool 28 // 29 //计算剩余的map和reduce的工作量:remaining 30 int remainingReduceLoad = 0; 31 int remainingMapLoad = 0; 32 synchronized (jobQueue) { 33 for (JobInProgress job : jobQueue) { 34 if (job.getStatus().getRunState() == JobStatus.RUNNING) { 35 remainingMapLoad += (job.desiredMaps() - job.finishedMaps()); 36 if (job.scheduleReduces()) { 37 remainingReduceLoad += 38 (job.desiredReduces() - job.finishedReduces()); 39 } 40 } 41 } 42 } 43 44 // Compute the 'load factor' for maps and reduces 45 double mapLoadFactor = 0.0; 46 if (clusterMapCapacity > 0) { 47 mapLoadFactor = (double)remainingMapLoad / clusterMapCapacity; 48 } 49 double reduceLoadFactor = 0.0; 50 if (clusterReduceCapacity > 0) { 51 reduceLoadFactor = (double)remainingReduceLoad / clusterReduceCapacity; 52 } 53 54 // 55 // In the below steps, we allocate first map tasks (if appropriate), 56 // and then reduce tasks if appropriate. We go through all jobs 57 // in order of job arrival; jobs only get serviced if their 58 // predecessors are serviced, too. 59 // 60 61 // 62 // We assign tasks to the current taskTracker if the given machine 63 // has a workload that's less than the maximum load of that kind of 64 // task. 65 // However, if the cluster is close to getting loaded i.e. we don't 66 // have enough _padding_ for speculative executions etc., we only 67 // schedule the "highest priority" task i.e. the task from the job 68 // with the highest priority. 69 // 70 71 final int trackerCurrentMapCapacity = 72 Math.min((int)Math.ceil(mapLoadFactor * trackerMapCapacity), 73 trackerMapCapacity); 74 int availableMapSlots = trackerCurrentMapCapacity - trackerRunningMaps; 75 boolean exceededMapPadding = false; 76 if (availableMapSlots > 0) { 77 exceededMapPadding = 78 exceededPadding(true, clusterStatus, trackerMapCapacity); 79 } 80 int numLocalMaps = 0; 81 int numNonLocalMaps = 0; 82 scheduleMaps: 83 for (int i=0; i < availableMapSlots; ++i) { 84 synchronized (jobQueue) { 85 for (JobInProgress job : jobQueue) { 86 if (job.getStatus().getRunState() != JobStatus.RUNNING) { 87 continue; 88 } 89 90 Task t = null; 91 92 // Try to schedule a node-local or rack-local Map task 93 t = 94 job.obtainNewNodeOrRackLocalMapTask(taskTrackerStatus, 95 numTaskTrackers, taskTrackerManager.getNumberOfUniqueHosts()); 96 if (t != null) { 97 assignedTasks.add(t); 98 ++numLocalMaps; 99 100 // Don't assign map tasks to the hilt!101 // Leave some free slots in the cluster for future task-failures,102 // speculative tasks etc. beyond the highest priority job103 if (exceededMapPadding) {104 break scheduleMaps;105 }106 107 // Try all jobs again for the next Map task 108 break;109 }110 111 // Try to schedule a node-local or rack-local Map task112 //产生 Map 任务使用 JobInProgress 的obtainNewMapTask() 方法,113 //实质上最后调用了 JobInProgress 的 findNewMapTask() 访问nonRunningMapCache 。114 t = 115 job.obtainNewNonLocalMapTask(taskTrackerStatus, numTaskTrackers,116 taskTrackerManager.getNumberOfUniqueHosts());117 118 if (t != null) {119 assignedTasks.add(t);120 ++numNonLocalMaps;121 122 // We assign at most 1 off-switch or speculative task123 // This is to prevent TaskTrackers from stealing local-tasks124 // from other TaskTrackers.125 break scheduleMaps;126 }127 }128 }129 }130 int assignedMaps = assignedTasks.size();131 132 //133 // Same thing, but for reduce tasks134 // However we _never_ assign more than 1 reduce task per heartbeat135 分配完map task,再分配reduce task 136 final int trackerCurrentReduceCapacity = 137 Math.min((int)Math.ceil(reduceLoadFactor * trackerReduceCapacity), 138 trackerReduceCapacity);139 final int availableReduceSlots = 140 Math.min((trackerCurrentReduceCapacity - trackerRunningReduces), 1);141 boolean exceededReducePadding = false;142 if (availableReduceSlots > 0) {143 exceededReducePadding = exceededPadding(false, clusterStatus, 144 trackerReduceCapacity);145 synchronized (jobQueue) {146 for (JobInProgress job : jobQueue) {147 if (job.getStatus().getRunState() != JobStatus.RUNNING ||148 job.numReduceTasks == 0) {149 continue;150 }151 //使用JobInProgress.obtainNewReduceTask() 方法,152 //实质上最后调用了JobInProgress的 findNewReduceTask() 访问 nonRuningReduceCache153 Task t = 154 job.obtainNewReduceTask(taskTrackerStatus, numTaskTrackers, 155 taskTrackerManager.getNumberOfUniqueHosts()156 );157 if (t != null) {158 assignedTasks.add(t);159 break;160 }161 162 // Don't assign reduce tasks to the hilt!163 // Leave some free slots in the cluster for future task-failures,164 // speculative tasks etc. beyond the highest priority job165 if (exceededReducePadding) {166 break;167 }168 }169 }170 }171 172 if (LOG.isDebugEnabled()) {173 LOG.debug("Task assignments for " + taskTrackerStatus.getTrackerName() + " --> " +174 "[" + mapLoadFactor + ", " + trackerMapCapacity + ", " + 175 trackerCurrentMapCapacity + ", " + trackerRunningMaps + "] -> [" + 176 (trackerCurrentMapCapacity - trackerRunningMaps) + ", " +177 assignedMaps + " (" + numLocalMaps + ", " + numNonLocalMaps + 178 ")] [" + reduceLoadFactor + ", " + trackerReduceCapacity + ", " + 179 trackerCurrentReduceCapacity + "," + trackerRunningReduces + 180 "] -> [" + (trackerCurrentReduceCapacity - trackerRunningReduces) + 181 ", " + (assignedTasks.size()-assignedMaps) + "]");182 }183 184 return assignedTasks;185 }
View Code

  该方法会先获取集群的基本信息,容量,然后获取此tasktracker的基本信息(slots及正在运行的task数);然后计算所有运行中的job的剩余量的总和(remainingReduceLoad和remainingMapLoad);分别计算map和reduce的负载因子(都是两种类型的剩余占对应的最大容量比)mapLoadFactor、reduceLoadFactor;然后计算trackerCurrentMapCapacity当前容量这里会使得集群中的所有tasktracker的负载尽量平均,因为Math.min((int)Math.ceil(mapLoadFactor * trackerMapCapacity), trackerMapCapacity),mapLoadFactor * trackerMapCapacity会使得该节点当前map的容量和集群整体的负载相近;然后获取当前tasktracker可用的mapslot,该tasktracker超过集群目前的负载水平后就不分配task,否则会有空闲的slot等待分配task;然后为每个mapslot选择一个map task,选择的过程十分复杂,首先会遍历jobQueue中的每个处于非运行状态的JobInProgress,调JobInProgress.obtainNewNodeOrRackLocalMapTask方法获取基于节点本地或者机架本地的map task,obtainNewNodeOrRackLocalMapTask会通过调用findNewMapTask获取map数组中的索引值。

  (1)首先从失败task选取合适的task直接返回。findNewMapTask方法会先通过findTaskFromList方法从failedMaps获取合适的失败map并返回(返回条件是A、该tasktracker没运行过TaskInProgress;B、该TaskInProgress失败过的节点数不低于运行taskTracker的主机数,这两个满足一个即可),如果有合适的失败map task,则通过scheduleMap(tip)方法将其加入nonLocalRunningMaps(该task没有对应的分片信息)或者runningMapCache(每个分片的存储Node及其对应的maptask列表,还有Node的父节点Node及对应的maptask列表也要加入),然后返回给obtainNewNodeOrRackLocalMapTask这个maptask在map数组中的索引值,此时从失败的task中寻找合适的task并不考虑数据的本地性。

  final SortedSet<TaskInProgress> failedMaps是按照task attempt失败次数排序的TaskInProgress集合。

  Set<TaskInProgress> nonLocalRunningMaps是no-local且正在运行的TaskInProgress结合。

  Map<Node, Set<TaskInProgress>> runningMapCache是Node与运行的TaskInProgress集合映射关系,一个任务获得调度机会,其TaskInProgress便会添加进来。

  (2)如果没有合适的失败task,则获取当前tasktracker对应的Node,然后“从近到远一层一层地寻找,直到找到合适的TaskInProgress”(通过不断获取父Node)从nonRunningMapCache中获取此Node的所有map task列表,如果列表不为空则调用findTaskFromList方法从这个列表中获取合适的TaskInProgress,如果tip!=null 则调用scheduleMap(tip)(上面已经介绍),然后检查列表是否为空,为空则从nonRunningMapCache清除这个Node的所有信息,再返回给obtainNewNodeOrRackLocalMapTask这个maptask在map数组中的索引值,如果遍历拓扑最大层数还是没有合适的task,则返回给obtainNewNodeOrRackLocalMapTask一个值-1,这里说明如果方法findNewMapTask的参数maxCacheLevel大于0则是获取(node-local或者rack-local,后面的其他情况不予考虑),其实就是优先考虑tasktracker对应Node有分片信息的本地的map(是node-local),然后再考虑父Node(同一个机架rack-local)的,再其他的(跨机架off-switch,这点得看设置的网络深度,大于2才会考虑),这样由近及远的做法会使得减少数据的拷贝距离,降低网络开销。  

  Map<Node, List<TaskInProgress>> nonRunningMapCache是Node与未运行的TaskInProgress的集合映射关系,通过作业的InputFormat可直接获取。

  (3)然后获取cache大网络深度的Node;获取该tasktracker对应Node的最深父Node;剩下的和上面(2)中的类似,只不过这次找的跨机架(或者更高一级,主要看设置的网络深度)。选择跨机架的task,scheduleMap(tip);返回给obtainNewNodeOrRackLocalMapTask这个maptask在map数组中的索引值。  

  (4)然后是查找nonLocalMaps中有无合适的task,这种任务没有输入数据,不需考虑本地性。scheduleMap(tip);返回给obtainNewNodeOrRackLocalMapTask这个maptask在map数组中的索引值。

  final List<TaskInProgress> nonLocalMaps是一些计算密集型任务,比如hadoop example中的PI作业。

  (5)如果有“拖后腿”的task(hasSpeculativeMaps==true),遍历runningMapCache,异常从node-local、rack-local、off-switch选择合适的“拖后腿”task,返回给obtainNewNodeOrRackLocalMapTask这个maptask在map数组中的索引值,这不需要scheduleMap(tip),很明显已经在runningMapCache中了。

  (6)从nonLocalRunningMaps中查找“拖后腿”的task,这是计算密集型任务在拖后腿,返回给obtainNewNodeOrRackLocalMapTask这个maptask在map数组中的索引值。

  (7)再找不到返回-1.

  obtainNewNodeOrRackLocalMapTask方法只执行到(2),要么返回一个MapTask要么返回null(findNewMapTask返回的是-1)这个maptask在map数组中的索引值,不再进行后续步骤。

  返回到obtainNewMapTask方法,获得map数组索引值后,还要获取该TaskInProgress的task(可能是MapTask或者ReduceTask,这里是MapTask),把这个task返回给assignTasks方法,加入分配task列表assignedTasks,跳出内层for循环,准备为下一个mapslot找合适的MapTask,如果没有合适的MapTask(node-local或者rack-local),则调用obtainNewNonLocalMapTask获取(除了上面的(2)不执行,其他都执行)MapTask,加入分配task列表assignedTasks,跳出内层for循环。

  然后分配ReduceTask,每次心跳分配不超过1个ReduceTask。和分配mapslot类似,这里至多分配一个reduceslot,遍历jobQueue通过obtainNewReduceTask方法获取合适的ReduceTask。obtainNewReduceTask方法会先做一个检查,和Map Task一样,会对节点的可靠性和磁盘空间进行检查;然后判断Job的map是否运行到该调用reduce的比例,若不到就返回null;然后调用findNewReduceTask方法获取reduce的索引值。findNewReduceTask方法会先检查该Job是否有reduce,没有就返回-1,检查此taskTracker是否可以运行reduce任务,然后调用方法findTaskFromList从nonRunningReduces中选择合适的TaskInProgress,放入runningReduces中,直接返回给obtainNewReduceTask对应的索引;如果没有合适的就从“拖后腿”的runningReduces中通过findSpeculativeTask方法找出退后退的reduce,放入runningReduces中,直接返回给obtainNewReduceTask对应的索引;再找不到就直接返回给obtainNewReduceTask方法-1。然后返回到obtainNewReduceTask方法,获取相应的ReduceTask,返回给assignTasks方法,加入分配任务列表assignedTasks中。

  在分配mapslot和reduceslot时循环中都有判断exceededReducePadding真假值的代码,exceededReducePadding是通过exceededPadding方法来获取的。在任务调度器JobQueueTaskScheduler的实现中,如果在集群中的TaskTracker节点比较多的情况下,它总是会想办法让若干个TaskTracker节点预留一些空闲的slots(计算能力),以便能够快速的处理优先级比较高的Job的Task或者发生错误的Task,以保证已经被调度的作业的完成。exceededPadding方法判断当前集群是否需要预留一部分map/reduce计算能力来执行那些失败的、紧急的或特殊的任务。

  还有一点需要注意的是对于每个slot总是会优先考虑jobQueue中的第一个job的任务(map、reduce),如果分配不成功才会考虑其他Job的,这样尽量保证优先处理第一个Job。

  assignTasks方法最后返回分配任务列表assignedTasks。调度器只分配MapTask和ReduceTask。而作业的其它辅助任务都是交由JobTracker来调度的,如JobSetup、JobCleanup、TaskCleanup任务等。

  对于JobQueueTaskScheduler的任务调度实现原则可总结如下:

     1.先调度优先级高的作业,统一优先级的作业则先进先出;
     2.尽量使集群每一个TaskTracker达到负载均衡(这个均衡是task数量上的而不是实际的工作强度);
     3.尽量分配作业的本地任务给TaskTracker,但不是尽快分配作业的本地任务给TaskTracker,最多分配一个非本地任务给TaskTracker(一是保证任务的并发性,二是避免有些TaskTracker的本地任务被偷走),最多分配一个reduce任务;
      4.为优先级或者紧急的Task预留一定的slot;

  七、遍历任务列表tasks,将所有task放入expireLaunchingTasks中监控是否过期expireLaunchingTasks.addNewTask(task.getTaskID()),然后放入actions.add(new LaunchTaskAction(task))。

  八、遍历taskTracker对应的所有task是否有需要kill的,以及trackerToTasksToCleanup中对应此tasktracker的task需要清理,封装成KillTaskAction,加入actions中。

  九、获取trackerToJobsToCleanup中对应此tasktracker的所有jobs,封装成KillJobAction,加入actions中。

  十、检查tasktracker的所有的task中状态等于TaskStatus.State.COMMIT_PENDING的,封装成CommitTaskAction,加入actions中。表示这个task的输出可以保存。

  十一、计算下一次心跳间隔与actions一同加入响应信息response。

  十二、如果JobTracker重启了,则将需要将需要恢复的Job列表加入response。response.setRecoveredJobs(recoveryManager.getJobsToRecover())

  十三、将trackerName及其响应信息response,加入trackerToHeartbeatResponseMap

  十四、因为已经将任务分配出去了,所以需要更新JobTracker的一些数据结构。removeMarkedTasks(trackerName)从一些相关的数据结构中清除trackerName对应的数据,比如trackerToMarkedTasksMap、taskidToTrackerMap、trackerToTaskMap、taskidToTIPMap等。

  十五、最后返回响应信息response。

 

参考:

1,、董西成,《hadoop技术内幕---深入理解MapReduce架构设计与实现原理》

2、http://blog.csdn.net/xhh198781/article/details/7046389

转载于:https://www.cnblogs.com/lxf20061900/p/3775963.html

你可能感兴趣的文章
浅谈算法和数据结构: 一 栈和队列
查看>>
Java内部类详解
查看>>
17 案例
查看>>
【hdu 1429】胜利大逃亡(续)
查看>>
图论-次短路求法
查看>>
What's New for Visual C# 6.0
查看>>
ExtJs学习笔记之ComboBox组件
查看>>
关于收费软件
查看>>
getopt_long
查看>>
TensorFlow MNIST CNN 代码
查看>>
javascript之Style物
查看>>
JSON跨域解决方案收集
查看>>
SSH框架整合总结
查看>>
图的深度优先遍历
查看>>
C# 之 提高WebService性能大数据量网络传输处理
查看>>
md5sum命令详解
查看>>
[bzoj1004] [HNOI2008] Cards
查看>>
应该是实例化对象的没有对属性赋值时,自动赋值为null,但不是空指针对象引用...
查看>>
原生HttpClient详细使用示例
查看>>
几道面试题
查看>>