Examples with Speculator org.apache.hadoop.mapreduce.v2.app.speculate.Speculator used on opensource projects

Example 1 with Speculator

use of org.apache.hadoop.mapreduce.v2.app.speculate.Speculator in project hadoop by apache.

the class MRAppMaster method serviceInit.

@Override
protected void serviceInit(final Configuration conf) throws Exception {
    // create the job classloader if enabled
    createJobClassLoader(conf);
    conf.setBoolean(Dispatcher.DISPATCHER_EXIT_ON_ERROR_KEY, true);
    initJobCredentialsAndUGI(conf);
    dispatcher = createDispatcher();
    addIfService(dispatcher);
    taskAttemptFinishingMonitor = createTaskAttemptFinishingMonitor(dispatcher.getEventHandler());
    addIfService(taskAttemptFinishingMonitor);
    context = new RunningAppContext(conf, taskAttemptFinishingMonitor);
    // Job name is the same as the app name util we support DAG of jobs
    // for an app later
    appName = conf.get(MRJobConfig.JOB_NAME, "<missing app name>");
    conf.setInt(MRJobConfig.APPLICATION_ATTEMPT_ID, appAttemptID.getAttemptId());
    newApiCommitter = false;
    jobId = MRBuilderUtils.newJobId(appAttemptID.getApplicationId(), appAttemptID.getApplicationId().getId());
    int numReduceTasks = conf.getInt(MRJobConfig.NUM_REDUCES, 0);
    if ((numReduceTasks > 0 && conf.getBoolean("mapred.reducer.new-api", false)) || (numReduceTasks == 0 && conf.getBoolean("mapred.mapper.new-api", false))) {
        newApiCommitter = true;
        LOG.info("Using mapred newApiCommitter.");
    }
    boolean copyHistory = false;
    committer = createOutputCommitter(conf);
    try {
        String user = UserGroupInformation.getCurrentUser().getShortUserName();
        Path stagingDir = MRApps.getStagingAreaDir(conf, user);
        FileSystem fs = getFileSystem(conf);
        boolean stagingExists = fs.exists(stagingDir);
        Path startCommitFile = MRApps.getStartJobCommitFile(conf, user, jobId);
        boolean commitStarted = fs.exists(startCommitFile);
        Path endCommitSuccessFile = MRApps.getEndJobCommitSuccessFile(conf, user, jobId);
        boolean commitSuccess = fs.exists(endCommitSuccessFile);
        Path endCommitFailureFile = MRApps.getEndJobCommitFailureFile(conf, user, jobId);
        boolean commitFailure = fs.exists(endCommitFailureFile);
        if (!stagingExists) {
            isLastAMRetry = true;
            LOG.info("Attempt num: " + appAttemptID.getAttemptId() + " is last retry: " + isLastAMRetry + " because the staging dir doesn't exist.");
            errorHappenedShutDown = true;
            forcedState = JobStateInternal.ERROR;
            shutDownMessage = "Staging dir does not exist " + stagingDir;
            LOG.fatal(shutDownMessage);
        } else if (commitStarted) {
            //A commit was started so this is the last time, we just need to know
            // what result we will use to notify, and how we will unregister
            errorHappenedShutDown = true;
            isLastAMRetry = true;
            LOG.info("Attempt num: " + appAttemptID.getAttemptId() + " is last retry: " + isLastAMRetry + " because a commit was started.");
            copyHistory = true;
            if (commitSuccess) {
                shutDownMessage = "Job commit succeeded in a prior MRAppMaster attempt " + "before it crashed. Recovering.";
                forcedState = JobStateInternal.SUCCEEDED;
            } else if (commitFailure) {
                shutDownMessage = "Job commit failed in a prior MRAppMaster attempt " + "before it crashed. Not retrying.";
                forcedState = JobStateInternal.FAILED;
            } else {
                if (isCommitJobRepeatable()) {
                    // cleanup previous half done commits if committer supports
                    // repeatable job commit.
                    errorHappenedShutDown = false;
                    cleanupInterruptedCommit(conf, fs, startCommitFile);
                } else {
                    //The commit is still pending, commit error
                    shutDownMessage = "Job commit from a prior MRAppMaster attempt is " + "potentially in progress. Preventing multiple commit executions";
                    forcedState = JobStateInternal.ERROR;
                }
            }
        }
    } catch (IOException e) {
        throw new YarnRuntimeException("Error while initializing", e);
    }
    if (errorHappenedShutDown) {
        NoopEventHandler eater = new NoopEventHandler();
        //We do not have a JobEventDispatcher in this path
        dispatcher.register(JobEventType.class, eater);
        EventHandler<JobHistoryEvent> historyService = null;
        if (copyHistory) {
            historyService = createJobHistoryHandler(context);
            dispatcher.register(org.apache.hadoop.mapreduce.jobhistory.EventType.class, historyService);
        } else {
            dispatcher.register(org.apache.hadoop.mapreduce.jobhistory.EventType.class, eater);
        }
        if (copyHistory) {
            // Now that there's a FINISHING state for application on RM to give AMs
            // plenty of time to clean up after unregister it's safe to clean staging
            // directory after unregistering with RM. So, we start the staging-dir
            // cleaner BEFORE the ContainerAllocator so that on shut-down,
            // ContainerAllocator unregisters first and then the staging-dir cleaner
            // deletes staging directory.
            addService(createStagingDirCleaningService());
        }
        // service to allocate containers from RM (if non-uber) or to fake it (uber)
        containerAllocator = createContainerAllocator(null, context);
        addIfService(containerAllocator);
        dispatcher.register(ContainerAllocator.EventType.class, containerAllocator);
        if (copyHistory) {
            // Add the JobHistoryEventHandler last so that it is properly stopped first.
            // This will guarantee that all history-events are flushed before AM goes
            // ahead with shutdown.
            // Note: Even though JobHistoryEventHandler is started last, if any
            // component creates a JobHistoryEvent in the meanwhile, it will be just be
            // queued inside the JobHistoryEventHandler 
            addIfService(historyService);
            JobHistoryCopyService cpHist = new JobHistoryCopyService(appAttemptID, dispatcher.getEventHandler());
            addIfService(cpHist);
        }
    } else {
        //service to handle requests from JobClient
        clientService = createClientService(context);
        // Init ClientService separately so that we stop it separately, since this
        // service needs to wait some time before it stops so clients can know the
        // final states
        clientService.init(conf);
        containerAllocator = createContainerAllocator(clientService, context);
        //service to handle the output committer
        committerEventHandler = createCommitterEventHandler(context, committer);
        addIfService(committerEventHandler);
        //policy handling preemption requests from RM
        callWithJobClassLoader(conf, new Action<Void>() {

            public Void call(Configuration conf) {
                preemptionPolicy = createPreemptionPolicy(conf);
                preemptionPolicy.init(context);
                return null;
            }
        });
        //service to handle requests to TaskUmbilicalProtocol
        taskAttemptListener = createTaskAttemptListener(context, preemptionPolicy);
        addIfService(taskAttemptListener);
        //service to log job history events
        EventHandler<JobHistoryEvent> historyService = createJobHistoryHandler(context);
        dispatcher.register(org.apache.hadoop.mapreduce.jobhistory.EventType.class, historyService);
        this.jobEventDispatcher = new JobEventDispatcher();
        //register the event dispatchers
        dispatcher.register(JobEventType.class, jobEventDispatcher);
        dispatcher.register(TaskEventType.class, new TaskEventDispatcher());
        dispatcher.register(TaskAttemptEventType.class, new TaskAttemptEventDispatcher());
        dispatcher.register(CommitterEventType.class, committerEventHandler);
        if (conf.getBoolean(MRJobConfig.MAP_SPECULATIVE, false) || conf.getBoolean(MRJobConfig.REDUCE_SPECULATIVE, false)) {
            //optional service to speculate on task attempts' progress
            speculator = createSpeculator(conf, context);
            addIfService(speculator);
        }
        speculatorEventDispatcher = new SpeculatorEventDispatcher(conf);
        dispatcher.register(Speculator.EventType.class, speculatorEventDispatcher);
        // Now that there's a FINISHING state for application on RM to give AMs
        // plenty of time to clean up after unregister it's safe to clean staging
        // directory after unregistering with RM. So, we start the staging-dir
        // cleaner BEFORE the ContainerAllocator so that on shut-down,
        // ContainerAllocator unregisters first and then the staging-dir cleaner
        // deletes staging directory.
        addService(createStagingDirCleaningService());
        // service to allocate containers from RM (if non-uber) or to fake it (uber)
        addIfService(containerAllocator);
        dispatcher.register(ContainerAllocator.EventType.class, containerAllocator);
        // corresponding service to launch allocated containers via NodeManager
        containerLauncher = createContainerLauncher(context);
        addIfService(containerLauncher);
        dispatcher.register(ContainerLauncher.EventType.class, containerLauncher);
        // Add the JobHistoryEventHandler last so that it is properly stopped first.
        // This will guarantee that all history-events are flushed before AM goes
        // ahead with shutdown.
        // Note: Even though JobHistoryEventHandler is started last, if any
        // component creates a JobHistoryEvent in the meanwhile, it will be just be
        // queued inside the JobHistoryEventHandler 
        addIfService(historyService);
    }
    super.serviceInit(conf);
}

Example 2 with Speculator

use of org.apache.hadoop.mapreduce.v2.app.speculate.Speculator in project hadoop by apache.

the class MRAppMaster method serviceStart.

@SuppressWarnings("unchecked")
@Override
protected void serviceStart() throws Exception {
    amInfos = new LinkedList<AMInfo>();
    completedTasksFromPreviousRun = new HashMap<TaskId, TaskInfo>();
    processRecovery();
    // Current an AMInfo for the current AM generation.
    AMInfo amInfo = MRBuilderUtils.newAMInfo(appAttemptID, startTime, containerID, nmHost, nmPort, nmHttpPort);
    // /////////////////// Create the job itself.
    job = createJob(getConfig(), forcedState, shutDownMessage);
    // Send out an MR AM inited event for all previous AMs.
    for (AMInfo info : amInfos) {
        dispatcher.getEventHandler().handle(new JobHistoryEvent(job.getID(), new AMStartedEvent(info.getAppAttemptId(), info.getStartTime(), info.getContainerId(), info.getNodeManagerHost(), info.getNodeManagerPort(), info.getNodeManagerHttpPort(), appSubmitTime)));
    }
    // Send out an MR AM inited event for this AM.
    dispatcher.getEventHandler().handle(new JobHistoryEvent(job.getID(), new AMStartedEvent(amInfo.getAppAttemptId(), amInfo.getStartTime(), amInfo.getContainerId(), amInfo.getNodeManagerHost(), amInfo.getNodeManagerPort(), amInfo.getNodeManagerHttpPort(), this.forcedState == null ? null : this.forcedState.toString(), appSubmitTime)));
    amInfos.add(amInfo);
    // metrics system init is really init & start.
    // It's more test friendly to put it here.
    DefaultMetricsSystem.initialize("MRAppMaster");
    boolean initFailed = false;
    if (!errorHappenedShutDown) {
        // create a job event for job intialization
        JobEvent initJobEvent = new JobEvent(job.getID(), JobEventType.JOB_INIT);
        // Send init to the job (this does NOT trigger job execution)
        // This is a synchronous call, not an event through dispatcher. We want
        // job-init to be done completely here.
        jobEventDispatcher.handle(initJobEvent);
        // If job is still not initialized, an error happened during
        // initialization. Must complete starting all of the services so failure
        // events can be processed.
        initFailed = (((JobImpl) job).getInternalState() != JobStateInternal.INITED);
        if (job.isUber()) {
            speculatorEventDispatcher.disableSpeculation();
            LOG.info("MRAppMaster uberizing job " + job.getID() + " in local container (\"uber-AM\") on node " + nmHost + ":" + nmPort + ".");
        } else {
            // send init to speculator only for non-uber jobs. 
            // This won't yet start as dispatcher isn't started yet.
            dispatcher.getEventHandler().handle(new SpeculatorEvent(job.getID(), clock.getTime()));
            LOG.info("MRAppMaster launching normal, non-uberized, multi-container " + "job " + job.getID() + ".");
        }
        // Start ClientService here, since it's not initialized if
        // errorHappenedShutDown is true
        clientService.start();
    }
    //start all the components
    super.serviceStart();
    // finally set the job classloader
    MRApps.setClassLoader(jobClassLoader, getConfig());
    // set job classloader if configured
    Limits.init(getConfig());
    if (initFailed) {
        JobEvent initFailedEvent = new JobEvent(job.getID(), JobEventType.JOB_INIT_FAILED);
        jobEventDispatcher.handle(initFailedEvent);
    } else {
        // All components have started, start the job.
        startJobs();
    }
}

Example 3 with Speculator

use of org.apache.hadoop.mapreduce.v2.app.speculate.Speculator in project hadoop by apache.

the class MRApp method submit.

public Job submit(Configuration conf, boolean mapSpeculative, boolean reduceSpeculative) throws Exception {
    String user = conf.get(MRJobConfig.USER_NAME, UserGroupInformation.getCurrentUser().getShortUserName());
    conf.set(MRJobConfig.USER_NAME, user);
    conf.set(MRJobConfig.MR_AM_STAGING_DIR, testAbsPath.toString());
    conf.setBoolean(MRJobConfig.MR_AM_CREATE_JH_INTERMEDIATE_BASE_DIR, true);
    // TODO: fix the bug where the speculator gets events with
    // not-fully-constructed objects. For now, disable speculative exec
    conf.setBoolean(MRJobConfig.MAP_SPECULATIVE, mapSpeculative);
    conf.setBoolean(MRJobConfig.REDUCE_SPECULATIVE, reduceSpeculative);
    init(conf);
    start();
    DefaultMetricsSystem.shutdown();
    Job job = getContext().getAllJobs().values().iterator().next();
    if (assignedQueue != null) {
        job.setQueueName(assignedQueue);
    }
    // Write job.xml
    String jobFile = MRApps.getJobFile(conf, user, TypeConverter.fromYarn(job.getID()));
    LOG.info("Writing job conf to " + jobFile);
    new File(jobFile).getParentFile().mkdirs();
    conf.writeXml(new FileOutputStream(jobFile));
    return job;
}

Example 4 with Speculator

use of org.apache.hadoop.mapreduce.v2.app.speculate.Speculator in project hadoop by apache.

the class TestRuntimeEstimators method coreTestEstimator.

private void coreTestEstimator(TaskRuntimeEstimator testedEstimator, int expectedSpeculations) {
    estimator = testedEstimator;
    clock = new ControlledClock();
    dispatcher = new AsyncDispatcher();
    myJob = null;
    slotsInUse.set(0);
    completedMaps.set(0);
    completedReduces.set(0);
    successfulSpeculations.set(0);
    taskTimeSavedBySpeculation.set(0);
    clock.tickMsec(1000);
    Configuration conf = new Configuration();
    myAppContext = new MyAppContext(MAP_TASKS, REDUCE_TASKS);
    myJob = myAppContext.getAllJobs().values().iterator().next();
    estimator.contextualize(conf, myAppContext);
    conf.setLong(MRJobConfig.SPECULATIVE_RETRY_AFTER_NO_SPECULATE, 500L);
    conf.setLong(MRJobConfig.SPECULATIVE_RETRY_AFTER_SPECULATE, 5000L);
    conf.setDouble(MRJobConfig.SPECULATIVECAP_RUNNING_TASKS, 0.1);
    conf.setDouble(MRJobConfig.SPECULATIVECAP_TOTAL_TASKS, 0.001);
    conf.setInt(MRJobConfig.SPECULATIVE_MINIMUM_ALLOWED_TASKS, 5);
    speculator = new DefaultSpeculator(conf, myAppContext, estimator, clock);
    Assert.assertEquals("wrong SPECULATIVE_RETRY_AFTER_NO_SPECULATE value", 500L, speculator.getSoonestRetryAfterNoSpeculate());
    Assert.assertEquals("wrong SPECULATIVE_RETRY_AFTER_SPECULATE value", 5000L, speculator.getSoonestRetryAfterSpeculate());
    Assert.assertEquals(speculator.getProportionRunningTasksSpeculatable(), 0.1, 0.00001);
    Assert.assertEquals(speculator.getProportionTotalTasksSpeculatable(), 0.001, 0.00001);
    Assert.assertEquals("wrong SPECULATIVE_MINIMUM_ALLOWED_TASKS value", 5, speculator.getMinimumAllowedSpeculativeTasks());
    dispatcher.register(Speculator.EventType.class, speculator);
    dispatcher.register(TaskEventType.class, new SpeculationRequestEventHandler());
    dispatcher.init(conf);
    dispatcher.start();
    speculator.init(conf);
    speculator.start();
    // Now that the plumbing is hooked up, we do the following:
    //  do until all tasks are finished, ...
    //  1: If we have spare capacity, assign as many map tasks as we can, then
    //     assign as many reduce tasks as we can.  Note that an odd reduce
    //     task might be started while there are still map tasks, because
    //     map tasks take 3 slots and reduce tasks 2 slots.
    //  2: Send a speculation event for every task attempt that's running
    //  note that new attempts might get started by the speculator
    // discover undone tasks
    int undoneMaps = MAP_TASKS;
    int undoneReduces = REDUCE_TASKS;
    // build a task sequence where all the maps precede any of the reduces
    List<Task> allTasksSequence = new LinkedList<Task>();
    allTasksSequence.addAll(myJob.getTasks(TaskType.MAP).values());
    allTasksSequence.addAll(myJob.getTasks(TaskType.REDUCE).values());
    while (undoneMaps + undoneReduces > 0) {
        undoneMaps = 0;
        undoneReduces = 0;
        // start all attempts which are new but for which there is enough slots
        for (Task task : allTasksSequence) {
            if (!task.isFinished()) {
                if (task.getType() == TaskType.MAP) {
                    ++undoneMaps;
                } else {
                    ++undoneReduces;
                }
            }
            for (TaskAttempt attempt : task.getAttempts().values()) {
                if (attempt.getState() == TaskAttemptState.NEW && INITIAL_NUMBER_FREE_SLOTS - slotsInUse.get() >= taskTypeSlots(task.getType())) {
                    MyTaskAttemptImpl attemptImpl = (MyTaskAttemptImpl) attempt;
                    SpeculatorEvent event = new SpeculatorEvent(attempt.getID(), false, clock.getTime());
                    speculator.handle(event);
                    attemptImpl.startUp();
                } else {
                    // If a task attempt is in progress we should send the news to
                    // the Speculator.
                    TaskAttemptStatus status = new TaskAttemptStatus();
                    status.id = attempt.getID();
                    status.progress = attempt.getProgress();
                    status.stateString = attempt.getState().name();
                    status.taskState = attempt.getState();
                    SpeculatorEvent event = new SpeculatorEvent(status, clock.getTime());
                    speculator.handle(event);
                }
            }
        }
        long startTime = System.currentTimeMillis();
        // drain the speculator event queue
        while (!speculator.eventQueueEmpty()) {
            Thread.yield();
            if (System.currentTimeMillis() > startTime + 130000) {
                return;
            }
        }
        clock.tickMsec(1000L);
        if (clock.getTime() % 10000L == 0L) {
            speculator.scanForSpeculations();
        }
    }
    Assert.assertEquals("We got the wrong number of successful speculations.", expectedSpeculations, successfulSpeculations.get());
}