* Licensed to the Apache Software Foundation (ASF) under one

* or more contributor license agreements.  See the NOTICE file

* distributed with this work for additional information

* regarding copyright ownership.  The ASF licenses this file

* to you under the Apache License, Version 2.0 (the

* "License"); you may not use this file except in compliance

* with the License.  You may obtain a copy of the License at

*

*     http://www.apache.org/licenses/LICENSE-2.0

*

* Unless required by applicable law or agreed to in writing, software

* distributed under the License is distributed on an "AS IS" BASIS,

* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

* See the License for the specific language governing permissions and

* limitations under the License.

*/

* A scheduler that schedules resources between a set of queues. The scheduler

* keeps track of the resources used by each queue, and attempts to maintain

* fairness by scheduling tasks at queues whose allocations are farthest below

* an ideal fair distribution.

*

* The fair scheduler supports hierarchical queues. All queues descend from a

* queue named "root". Available resources are distributed among the children

* of the root queue in the typical fair scheduling fashion. Then, the children

* distribute the resources assigned to them to their children in the same

* fashion.  Applications may only be scheduled on leaf queues. Queues can be

* specified as children of other queues by placing them as sub-elements of their

* parents in the fair scheduler configuration file.

*

* A queue's name starts with the names of its parents, with periods as

* separators.  So a queue named "queue1" under the root named, would be

* referred to as "root.queue1", and a queue named "queue2" under a queue

* named "parent1" would be referred to as "root.parent1.queue2".

*/

@SuppressWarnings("unchecked")

  private boolean initialized;

  private FairSchedulerConfiguration conf;

  private Resource minimumAllocation;

  private Resource maximumAllocation;

  private Resource incrAllocation;

  private QueueManager queueMgr;

  private Clock clock;

  private boolean usePortForNodeName;

  private static final Log LOG = LogFactory.getLog(FairScheduler.class);

 

  private static final ResourceCalculator RESOURCE_CALCULATOR =

      new DefaultResourceCalculator();

 

  // Value that container assignment methods return when a container is

  // reserved

  public static final Resource CONTAINER_RESERVED = Resources.createResource(-1);

  // How often fair shares are re-calculated (ms)

  protected long UPDATE_INTERVAL = 500;

  // Aggregate metrics

  FSQueueMetrics rootMetrics;

  // Time when we last updated preemption vars

  protected long lastPreemptionUpdateTime;

  // Time we last ran preemptTasksIfNecessary

  private long lastPreemptCheckTime;

  // Nodes in the cluster, indexed by NodeId

  private Map<NodeId, FSSchedulerNode> nodes =

      new ConcurrentHashMap<NodeId, FSSchedulerNode>();

  // Aggregate capacity of the cluster

  private Resource clusterCapacity =

      RecordFactoryProvider.getRecordFactory(null).newRecordInstance(Resource.class);

  // How often tasks are preempted

  protected long preemptionInterval;

 

  // ms to wait before force killing stuff (must be longer than a couple

  // of heartbeats to give task-kill commands a chance to act).

  protected long waitTimeBeforeKill;

 

  // Containers whose AMs have been warned that they will be preempted soon.

  private List<RMContainer> warnedContainers = new ArrayList<RMContainer>();

 

  protected boolean preemptionEnabled;

  protected boolean sizeBasedWeight; // Give larger weights to larger jobs

  protected WeightAdjuster weightAdjuster; // Can be null for no weight adjuster

  protected boolean continuousSchedulingEnabled; // Continuous Scheduling enabled or not

  protected int continuousSchedulingSleepMs; // Sleep time for each pass in continuous scheduling

  private Comparator nodeAvailableResourceComparator =

          new NodeAvailableResourceComparator(); // Node available resource comparator

  protected double nodeLocalityThreshold; // Cluster threshold for node locality

  protected double rackLocalityThreshold; // Cluster threshold for rack locality

  protected long nodeLocalityDelayMs; // Delay for node locality

  protected long rackLocalityDelayMs; // Delay for rack locality

  private FairSchedulerEventLog eventLog; // Machine-readable event log

  protected boolean assignMultiple; // Allocate multiple containers per

                                    // heartbeat

  protected int maxAssign; // Max containers to assign per heartbeat

  @VisibleForTesting

  final MaxRunningAppsEnforcer maxRunningEnforcer;

  private AllocationFileLoaderService allocsLoader;

  @VisibleForTesting

  AllocationConfiguration allocConf;

 

  public FairScheduler() {

    clock = new SystemClock();

    allocsLoader = new AllocationFileLoaderService();

    queueMgr = new QueueManager(this);

    maxRunningEnforcer = new MaxRunningAppsEnforcer(this);

  }

  private void validateConf(Configuration conf) {

    // validate scheduler memory allocation setting

    int minMem = conf.getInt(

      YarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_MB,

      YarnConfiguration.DEFAULT_RM_SCHEDULER_MINIMUM_ALLOCATION_MB);

    int maxMem = conf.getInt(

      YarnConfiguration.RM_SCHEDULER_MAXIMUM_ALLOCATION_MB,

      YarnConfiguration.DEFAULT_RM_SCHEDULER_MAXIMUM_ALLOCATION_MB);

    if (minMem < 0 || minMem > maxMem) {

      throw new YarnRuntimeException("Invalid resource scheduler memory"

        + " allocation configuration"

        + ", " + YarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_MB

        + "=" + minMem

        + ", " + YarnConfiguration.RM_SCHEDULER_MAXIMUM_ALLOCATION_MB

        + "=" + maxMem + ", min should equal greater than 0"

        + ", max should be no smaller than min.");

    }

    // validate scheduler vcores allocation setting

    int minVcores = conf.getInt(

      YarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_VCORES,

      YarnConfiguration.DEFAULT_RM_SCHEDULER_MINIMUM_ALLOCATION_VCORES);

    int maxVcores = conf.getInt(

      YarnConfiguration.RM_SCHEDULER_MAXIMUM_ALLOCATION_VCORES,

      YarnConfiguration.DEFAULT_RM_SCHEDULER_MAXIMUM_ALLOCATION_VCORES);

    if (minVcores < 0 || minVcores > maxVcores) {

      throw new YarnRuntimeException("Invalid resource scheduler vcores"

        + " allocation configuration"

        + ", " + YarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_VCORES

        + "=" + minVcores

        + ", " + YarnConfiguration.RM_SCHEDULER_MAXIMUM_ALLOCATION_VCORES

        + "=" + maxVcores + ", min should equal greater than 0"

        + ", max should be no smaller than min.");

    }

  }

  public FairSchedulerConfiguration getConf() {

    return conf;

  }

  public QueueManager getQueueManager() {

    return queueMgr;

  }

  @Override

  public RMContainer getRMContainer(ContainerId containerId) {

    FSSchedulerApp attempt = getCurrentAttemptForContainer(containerId);

    return (attempt == null) ? null : attempt.getRMContainer(containerId);

  }

  private FSSchedulerApp getCurrentAttemptForContainer(

      ContainerId containerId) {

    SchedulerApplication app =

        applications.get(containerId.getApplicationAttemptId()

          .getApplicationId());

    if (app != null) {

      return (FSSchedulerApp) app.getCurrentAppAttempt();

    }

    return null;

  }

   * A runnable which calls {@link FairScheduler#update()} every

   * <code>UPDATE_INTERVAL</code> milliseconds.

   */

  private class UpdateThread implements Runnable {

    public void run() {

      while (true) {

        try {

          Thread.sleep(UPDATE_INTERVAL);

          update();

          preemptTasksIfNecessary();

        } catch (Exception e) {

          LOG.error("Exception in fair scheduler UpdateThread", e);

        }

      }

    }

  }

  /**

   * Recompute the internal variables used by the scheduler - per-job weights,

   * fair shares, deficits, minimum slot allocations, and amount of used and

   * required resources per job.

   */

  protected synchronized void update() {

    updatePreemptionVariables(); // Determine if any queues merit preemption

    FSQueue rootQueue = queueMgr.getRootQueue();

    // Recursively update demands for all queues

    rootQueue.updateDemand();

    rootQueue.setFairShare(clusterCapacity);

    // Recursively compute fair shares for all queues

    // and update metrics

    rootQueue.recomputeShares();

  }

  /**

   * Update the preemption fields for all QueueScheduables, i.e. the times since

   * each queue last was at its guaranteed share and at > 1/2 of its fair share

   * for each type of task.

   */

  private void updatePreemptionVariables() {

    long now = clock.getTime();

    lastPreemptionUpdateTime = now;

    for (FSLeafQueue sched : queueMgr.getLeafQueues()) {

      if (!isStarvedForMinShare(sched)) {

        sched.setLastTimeAtMinShare(now);

      }

      if (!isStarvedForFairShare(sched)) {

        sched.setLastTimeAtHalfFairShare(now);

      }

    }

  }

  /**

   * Is a queue below its min share for the given task type?

   */

  boolean isStarvedForMinShare(FSLeafQueue sched) {

    Resource desiredShare = Resources.min(RESOURCE_CALCULATOR, clusterCapacity,

      sched.getMinShare(), sched.getDemand());

    return Resources.lessThan(RESOURCE_CALCULATOR, clusterCapacity,

        sched.getResourceUsage(), desiredShare);

  }

  /**

   * Is a queue being starved for fair share for the given task type? This is

   * defined as being below half its fair share.

   */

  boolean isStarvedForFairShare(FSLeafQueue sched) {

    Resource desiredFairShare = Resources.min(RESOURCE_CALCULATOR, clusterCapacity,

        Resources.multiply(sched.getFairShare(), .5), sched.getDemand());

    return Resources.lessThan(RESOURCE_CALCULATOR, clusterCapacity,

        sched.getResourceUsage(), desiredFairShare);

  }

  /**

   * Check for queues that need tasks preempted, either because they have been

   * below their guaranteed share for minSharePreemptionTimeout or they have

   * been below half their fair share for the fairSharePreemptionTimeout. If

   * such queues exist, compute how many tasks of each type need to be preempted

   * and then select the right ones using preemptTasks.

   */

  protected synchronized void preemptTasksIfNecessary() {

    if (!preemptionEnabled) {

      return;

    }

    long curTime = clock.getTime();

    if (curTime - lastPreemptCheckTime < preemptionInterval) {

      return;

    }

    lastPreemptCheckTime = curTime;

    Resource resToPreempt = Resources.none();

    for (FSLeafQueue sched : queueMgr.getLeafQueues()) {

      resToPreempt = Resources.add(resToPreempt, resToPreempt(sched, curTime));

    }

    if (Resources.greaterThan(RESOURCE_CALCULATOR, clusterCapacity, resToPreempt,

        Resources.none())) {

      preemptResources(queueMgr.getLeafQueues(), resToPreempt);

    }

  }

  /**

   * Preempt a quantity of resources from a list of QueueSchedulables. The

   * policy for this is to pick apps from queues that are over their fair share,

   * but make sure that no queue is placed below its fair share in the process.

   * We further prioritize preemption by choosing containers with lowest

   * priority to preempt.

   */

  protected void preemptResources(Collection<FSLeafQueue> scheds,

      Resource toPreempt) {

    if (scheds.isEmpty() || Resources.equals(toPreempt, Resources.none())) {

      return;

    }

    Map<RMContainer, FSSchedulerApp> apps =

        new HashMap<RMContainer, FSSchedulerApp>();

    Map<RMContainer, FSLeafQueue> queues =

        new HashMap<RMContainer, FSLeafQueue>();

    // Collect running containers from over-scheduled queues

    List<RMContainer> runningContainers = new ArrayList<RMContainer>();

    for (FSLeafQueue sched : scheds) {

      if (Resources.greaterThan(RESOURCE_CALCULATOR, clusterCapacity,

          sched.getResourceUsage(), sched.getFairShare())) {

        for (AppSchedulable as : sched.getRunnableAppSchedulables()) {

          for (RMContainer c : as.getApp().getLiveContainers()) {

            runningContainers.add(c);

            apps.put(c, as.getApp());

            queues.put(c, sched);

          }

        }

      }

    }

    // Sort containers into reverse order of priority

    Collections.sort(runningContainers, new Comparator<RMContainer>() {

      public int compare(RMContainer c1, RMContainer c2) {

        int ret = c1.getContainer().getPriority().compareTo(

            c2.getContainer().getPriority());

        if (ret == 0) {

          return c2.getContainerId().compareTo(c1.getContainerId());

        }

        return ret;

      }

    });

   

    // Scan down the list of containers we've already warned and kill them

    // if we need to.  Remove any containers from the list that we don't need

    // or that are no longer running.

    Iterator<RMContainer> warnedIter = warnedContainers.iterator();

    Set<RMContainer> preemptedThisRound = new HashSet<RMContainer>();

    while (warnedIter.hasNext()) {

      RMContainer container = warnedIter.next();

      if (container.getState() == RMContainerState.RUNNING &&

          Resources.greaterThan(RESOURCE_CALCULATOR, clusterCapacity,

              toPreempt, Resources.none())) {

        warnOrKillContainer(container, apps.get(container), queues.get(container));

        preemptedThisRound.add(container);

        Resources.subtractFrom(toPreempt, container.getContainer().getResource());

      } else {

        warnedIter.remove();

      }

    }

    // Scan down the rest of the containers until we've preempted enough, making

    // sure we don't preempt too many from any queue

    Iterator<RMContainer> runningIter = runningContainers.iterator();

    while (runningIter.hasNext() &&

        Resources.greaterThan(RESOURCE_CALCULATOR, clusterCapacity,

            toPreempt, Resources.none())) {

      RMContainer container = runningIter.next();

      FSLeafQueue sched = queues.get(container);

      if (!preemptedThisRound.contains(container) &&

          Resources.greaterThan(RESOURCE_CALCULATOR, clusterCapacity,

              sched.getResourceUsage(), sched.getFairShare())) {

        warnOrKillContainer(container, apps.get(container), sched);

       

        warnedContainers.add(container);

        Resources.subtractFrom(toPreempt, container.getContainer().getResource());

      }

    }

  }

 

  private void warnOrKillContainer(RMContainer container, FSSchedulerApp app,

      FSLeafQueue queue) {

    LOG.info("Preempting container (prio=" + container.getContainer().getPriority() +

        "res=" + container.getContainer().getResource() +

        ") from queue " + queue.getName());

   

    Long time = app.getContainerPreemptionTime(container);

    if (time != null) {

      // if we asked for preemption more than maxWaitTimeBeforeKill ms ago,

      // proceed with kill

      if (time + waitTimeBeforeKill < clock.getTime()) {

        ContainerStatus status =

          SchedulerUtils.createPreemptedContainerStatus(

            container.getContainerId(), SchedulerUtils.PREEMPTED_CONTAINER);

        // TODO: Not sure if this ever actually adds this to the list of cleanup

        // containers on the RMNode (see SchedulerNode.releaseContainer()).

        completedContainer(container, status, RMContainerEventType.KILL);

        LOG.info("Killing container" + container +

            " (after waiting for premption for " +

            (clock.getTime() - time) + "ms)");

      }

    } else {

      // track the request in the FSSchedulerApp itself

      app.addPreemption(container, clock.getTime());

    }

  }

  /**

   * Return the resource amount that this queue is allowed to preempt, if any.

   * If the queue has been below its min share for at least its preemption

   * timeout, it should preempt the difference between its current share and

   * this min share. If it has been below half its fair share for at least the

   * fairSharePreemptionTimeout, it should preempt enough tasks to get up to its

   * full fair share. If both conditions hold, we preempt the max of the two

   * amounts (this shouldn't happen unless someone sets the timeouts to be

   * identical for some reason).

   */

  protected Resource resToPreempt(FSLeafQueue sched, long curTime) {

    String queue = sched.getName();

    long minShareTimeout = allocConf.getMinSharePreemptionTimeout(queue);

    long fairShareTimeout = allocConf.getFairSharePreemptionTimeout();

    Resource resDueToMinShare = Resources.none();

    Resource resDueToFairShare = Resources.none();

    if (curTime - sched.getLastTimeAtMinShare() > minShareTimeout) {

      Resource target = Resources.min(RESOURCE_CALCULATOR, clusterCapacity,

          sched.getMinShare(), sched.getDemand());

      resDueToMinShare = Resources.max(RESOURCE_CALCULATOR, clusterCapacity,

          Resources.none(), Resources.subtract(target, sched.getResourceUsage()));

    }

    if (curTime - sched.getLastTimeAtHalfFairShare() > fairShareTimeout) {

      Resource target = Resources.min(RESOURCE_CALCULATOR, clusterCapacity,

          sched.getFairShare(), sched.getDemand());

      resDueToFairShare = Resources.max(RESOURCE_CALCULATOR, clusterCapacity,

          Resources.none(), Resources.subtract(target, sched.getResourceUsage()));

    }

    Resource resToPreempt = Resources.max(RESOURCE_CALCULATOR, clusterCapacity,

        resDueToMinShare, resDueToFairShare);

    if (Resources.greaterThan(RESOURCE_CALCULATOR, clusterCapacity,

        resToPreempt, Resources.none())) {

      String message = "Should preempt " + resToPreempt + " res for queue "

          + sched.getName() + ": resDueToMinShare = " + resDueToMinShare

          + ", resDueToFairShare = " + resDueToFairShare;

      LOG.info(message);

    }

    return resToPreempt;

  }

  public RMContainerTokenSecretManager getContainerTokenSecretManager() {

    return rmContext.getContainerTokenSecretManager();

  }

  // synchronized for sizeBasedWeight

  public synchronized ResourceWeights getAppWeight(AppSchedulable app) {

    double weight = 1.0;

    if (sizeBasedWeight) {

      // Set weight based on current memory demand

      weight = Math.log1p(app.getDemand().getMemory()) / Math.log(2);

    }

    weight *= app.getPriority().getPriority();

    if (weightAdjuster != null) {

      // Run weight through the user-supplied weightAdjuster

      weight = weightAdjuster.adjustWeight(app, weight);

    }

    return new ResourceWeights((float)weight);

  }

  @Override

  public Resource getMinimumResourceCapability() {

    return minimumAllocation;

  }

  public Resource getIncrementResourceCapability() {

    return incrAllocation;

  }

  @Override

  public Resource getMaximumResourceCapability() {

    return maximumAllocation;

  }

  public double getNodeLocalityThreshold() {

    return nodeLocalityThreshold;

  }

  public double getRackLocalityThreshold() {

    return rackLocalityThreshold;

  }

  public long getNodeLocalityDelayMs() {

    return nodeLocalityDelayMs;

  }

  public long getRackLocalityDelayMs() {

    return rackLocalityDelayMs;

  }

  public boolean isContinuousSchedulingEnabled() {

    return continuousSchedulingEnabled;

  }

  public synchronized int getContinuousSchedulingSleepMs() {

    return continuousSchedulingSleepMs;

  }

  public Resource getClusterCapacity() {

    return clusterCapacity;

  }

  public synchronized Clock getClock() {

    return clock;

  }

  protected synchronized void setClock(Clock clock) {

    this.clock = clock;

  }

  public FairSchedulerEventLog getEventLog() {

    return eventLog;

  }

  /**

   * Add a new application to the scheduler, with a given id, queue name, and

   * user. This will accept a new app even if the user or queue is above

   * configured limits, but the app will not be marked as runnable.

   */

  protected synchronized void addApplication(ApplicationId applicationId,

      String queueName, String user) {

    if (queueName == null || queueName.isEmpty()) {

      String message = "Reject application " + applicationId +

              " submitted by user " + user + " with an empty queue name.";

      LOG.info(message);

      rmContext.getDispatcher().getEventHandler()

          .handle(new RMAppRejectedEvent(applicationId, message));

      return;

    }

    RMApp rmApp = rmContext.getRMApps().get(applicationId);

    FSLeafQueue queue = assignToQueue(rmApp, queueName, user);

    if (queue == null) {

      return;

    }

    // Enforce ACLs

    UserGroupInformation userUgi = UserGroupInformation.createRemoteUser(user);

    if (!queue.hasAccess(QueueACL.SUBMIT_APPLICATIONS, userUgi)

        && !queue.hasAccess(QueueACL.ADMINISTER_QUEUE, userUgi)) {

      String msg = "User " + userUgi.getUserName() +

              " cannot submit applications to queue " + queue.getName();

      LOG.info(msg);

      rmContext.getDispatcher().getEventHandler()

          .handle(new RMAppRejectedEvent(applicationId, msg));

      return;

    }

 

    SchedulerApplication application =

        new SchedulerApplication(queue, user);

    applications.put(applicationId, application);

    queue.getMetrics().submitApp(user);

    LOG.info("Accepted application " + applicationId + " from user: " + user

        + ", in queue: " + queueName + ", currently num of applications: "

        + applications.size());

    rmContext.getDispatcher().getEventHandler()

        .handle(new RMAppEvent(applicationId, RMAppEventType.APP_ACCEPTED));

  }

  /**

   * Add a new application attempt to the scheduler.

   */

  protected synchronized void addApplicationAttempt(

      ApplicationAttemptId applicationAttemptId,

      boolean transferStateFromPreviousAttempt) {

    SchedulerApplication application =

        applications.get(applicationAttemptId.getApplicationId());

    String user = application.getUser();

    FSLeafQueue queue = (FSLeafQueue) application.getQueue();

    FSSchedulerApp attempt =

        new FSSchedulerApp(applicationAttemptId, user,

            queue, new ActiveUsersManager(getRootQueueMetrics()),

            rmContext);

    if (transferStateFromPreviousAttempt) {

      attempt.transferStateFromPreviousAttempt(application

        .getCurrentAppAttempt());

    }

    application.setCurrentAppAttempt(attempt);

    boolean runnable = maxRunningEnforcer.canAppBeRunnable(queue, user);

    queue.addApp(attempt, runnable);

    if (runnable) {

      maxRunningEnforcer.trackRunnableApp(attempt);

    } else {

      maxRunningEnforcer.trackNonRunnableApp(attempt);

    }

   

    queue.getMetrics().submitAppAttempt(user);

    LOG.info("Added Application Attempt " + applicationAttemptId

        + " to scheduler from user: " + user);

    rmContext.getDispatcher().getEventHandler().handle(

        new RMAppAttemptEvent(applicationAttemptId,

            RMAppAttemptEventType.ATTEMPT_ADDED));

  }

  /**

   * Helper method that attempts to assign the app to a queue. The method is

   * responsible to call the appropriate event-handler if the app is rejected.

   */

  @VisibleForTesting

  FSLeafQueue assignToQueue(RMApp rmApp, String queueName, String user) {

    FSLeafQueue queue = null;

    String appRejectMsg = null;

    try {

      QueuePlacementPolicy placementPolicy = allocConf.getPlacementPolicy();

      queueName = placementPolicy.assignAppToQueue(queueName, user);

      if (queueName == null) {

        appRejectMsg = "Application rejected by queue placement policy";

      } else {

        queue = queueMgr.getLeafQueue(queueName, true);

        if (queue == null) {

          appRejectMsg = queueName + " is not a leaf queue";

        }

      }

    } catch (IOException ioe) {

      appRejectMsg = "Error assigning app to queue " + queueName;

    }

    if (appRejectMsg != null && rmApp != null) {

      LOG.error(appRejectMsg);

      rmContext.getDispatcher().getEventHandler().handle(

          new RMAppRejectedEvent(rmApp.getApplicationId(), appRejectMsg));

      return null;

    }

    if (rmApp != null) {

      rmApp.setQueue(queue.getName());

    } else {

      LOG.error("Couldn't find RM app to set queue name on");

    }

    return queue;

  }

  private synchronized void removeApplication(ApplicationId applicationId,

      RMAppState finalState) {

    SchedulerApplication application = applications.get(applicationId);

    if (application == null){

      LOG.warn("Couldn't find application " + applicationId);

      return;

    }

    application.stop(finalState);

    applications.remove(applicationId);

  }

  private synchronized void removeApplicationAttempt(

      ApplicationAttemptId applicationAttemptId,

      RMAppAttemptState rmAppAttemptFinalState, boolean keepContainers) {

    LOG.info("Application " + applicationAttemptId + " is done." +

        " finalState=" + rmAppAttemptFinalState);

    SchedulerApplication application =

        applications.get(applicationAttemptId.getApplicationId());

    FSSchedulerApp attempt = getSchedulerApp(applicationAttemptId);

    if (attempt == null || application == null) {

      LOG.info("Unknown application " + applicationAttemptId + " has completed!");

      return;

    }

    // Release all the running containers

    for (RMContainer rmContainer : attempt.getLiveContainers()) {

      if (keepContainers

          && rmContainer.getState().equals(RMContainerState.RUNNING)) {

        // do not kill the running container in the case of work-preserving AM

        // restart.

        LOG.info("Skip killing " + rmContainer.getContainerId());

        continue;

      }

      completedContainer(rmContainer,

          SchedulerUtils.createAbnormalContainerStatus(

              rmContainer.getContainerId(),

              SchedulerUtils.COMPLETED_APPLICATION),

              RMContainerEventType.KILL);

    }

    // Release all reserved containers

    for (RMContainer rmContainer : attempt.getReservedContainers()) {

      completedContainer(rmContainer,

          SchedulerUtils.createAbnormalContainerStatus(

              rmContainer.getContainerId(),

              "Application Complete"),

              RMContainerEventType.KILL);

    }

    // Clean up pending requests, metrics etc.

    attempt.stop(rmAppAttemptFinalState);

    // Inform the queue

    FSLeafQueue queue = queueMgr.getLeafQueue(attempt.getQueue()

        .getQueueName(), false);

    boolean wasRunnable = queue.removeApp(attempt);

    if (wasRunnable) {

      maxRunningEnforcer.untrackRunnableApp(attempt);

      maxRunningEnforcer.updateRunnabilityOnAppRemoval(attempt,

          attempt.getQueue());

    } else {

      maxRunningEnforcer.untrackNonRunnableApp(attempt);

    }

  }

  /**

   * Clean up a completed container.

   */

  private synchronized void completedContainer(RMContainer rmContainer,

      ContainerStatus containerStatus, RMContainerEventType event) {

    if (rmContainer == null) {

      LOG.info("Null container completed...");

      return;

    }

    Container container = rmContainer.getContainer();

    // Get the application for the finished container

    FSSchedulerApp application =

        getCurrentAttemptForContainer(container.getId());

    ApplicationId appId =

        container.getId().getApplicationAttemptId().getApplicationId();

    if (application == null) {

      LOG.info("Container " + container + " of" +

          " unknown application attempt " + appId +

          " completed with event " + event);

      return;

    }

    // Get the node on which the container was allocated

    FSSchedulerNode node = nodes.get(container.getNodeId());

    if (rmContainer.getState() == RMContainerState.RESERVED) {

      application.unreserve(node, rmContainer.getReservedPriority());

      node.unreserveResource(application);

    } else {

      application.containerCompleted(rmContainer, containerStatus, event);

      node.releaseContainer(container);

      updateRootQueueMetrics();

    }

    LOG.info("Application attempt " + application.getApplicationAttemptId()

        + " released container " + container.getId() + " on node: " + node

        + " with event: " + event);

  }

  private synchronized void addNode(RMNode node) {

    nodes.put(node.getNodeID(), new FSSchedulerNode(node, usePortForNodeName));

    Resources.addTo(clusterCapacity, node.getTotalCapability());

    updateRootQueueMetrics();

    LOG.info("Added node " + node.getNodeAddress() +

        " cluster capacity: " + clusterCapacity);

  }

  private synchronized void removeNode(RMNode rmNode) {

    FSSchedulerNode node = nodes.get(rmNode.getNodeID());

    // This can occur when an UNHEALTHY node reconnects

    if (node == null) {

      return;

    }

    Resources.subtractFrom(clusterCapacity, rmNode.getTotalCapability());

    updateRootQueueMetrics();

    // Remove running containers

    List<RMContainer> runningContainers = node.getRunningContainers();

    for (RMContainer container : runningContainers) {

      completedContainer(container,

          SchedulerUtils.createAbnormalContainerStatus(

              container.getContainerId(),

              SchedulerUtils.LOST_CONTAINER),

          RMContainerEventType.KILL);

    }

    // Remove reservations, if any

    RMContainer reservedContainer = node.getReservedContainer();

    if (reservedContainer != null) {

      completedContainer(reservedContainer,

          SchedulerUtils.createAbnormalContainerStatus(

              reservedContainer.getContainerId(),

              SchedulerUtils.LOST_CONTAINER),

          RMContainerEventType.KILL);

    }

    nodes.remove(rmNode.getNodeID());

    LOG.info("Removed node " + rmNode.getNodeAddress() +

        " cluster capacity: " + clusterCapacity);

  }

  @Override

  public Allocation allocate(ApplicationAttemptId appAttemptId,

      List<ResourceRequest> ask, List<ContainerId> release, List<String> blacklistAdditions, List<String> blacklistRemovals) {

    // Make sure this application exists

    FSSchedulerApp application = getSchedulerApp(appAttemptId);

    if (application == null) {

      LOG.info("Calling allocate on removed " +

          "or non existant application " + appAttemptId);

      return EMPTY_ALLOCATION;

    }

    // Sanity check

    SchedulerUtils.normalizeRequests(ask, new DominantResourceCalculator(),

        clusterCapacity, minimumAllocation, maximumAllocation, incrAllocation);

    // Release containers

    for (ContainerId releasedContainerId : release) {

      RMContainer rmContainer = getRMContainer(releasedContainerId);

      if (rmContainer == null) {

        RMAuditLogger.logFailure(application.getUser(),

            AuditConstants.RELEASE_CONTAINER,

            "Unauthorized access or invalid container", "FairScheduler",

            "Trying to release container not owned by app or with invalid id",

            application.getApplicationId(), releasedContainerId);

      }

      completedContainer(rmContainer,

          SchedulerUtils.createAbnormalContainerStatus(

              releasedContainerId,

              SchedulerUtils.RELEASED_CONTAINER),

          RMContainerEventType.RELEASED);

    }

    synchronized (application) {

      if (!ask.isEmpty()) {

        if (LOG.isDebugEnabled()) {

          LOG.debug("allocate: pre-update" +

              " applicationAttemptId=" + appAttemptId +

              " application=" + application.getApplicationId());

        }

        application.showRequests();

        // Update application requests

        application.updateResourceRequests(ask);

        LOG.debug("allocate: post-update");

        application.showRequests();

      }

      if (LOG.isDebugEnabled()) {

        LOG.debug("allocate:" +

            " applicationAttemptId=" + appAttemptId +

            " #ask=" + ask.size());

        LOG.debug("Preempting " + application.getPreemptionContainers().size()

            + " container(s)");

      }

     

      Set<ContainerId> preemptionContainerIds = new HashSet<ContainerId>();

      for (RMContainer container : application.getPreemptionContainers()) {

        preemptionContainerIds.add(container.getContainerId());

      }

      application.updateBlacklist(blacklistAdditions, blacklistRemovals);

      ContainersAndNMTokensAllocation allocation =

          application.pullNewlyAllocatedContainersAndNMTokens();

      return new Allocation(allocation.getContainerList(),

        application.getHeadroom(), preemptionContainerIds, null, null,

        allocation.getNMTokenList());

    }

  }

  /**

   * Process a container which has launched on a node, as reported by the node.

   */

  private void containerLaunchedOnNode(ContainerId containerId, FSSchedulerNode node) {

    // Get the application for the finished container

    FSSchedulerApp application = getCurrentAttemptForContainer(containerId);

    if (application == null) {

      LOG.info("Unknown application "

          + containerId.getApplicationAttemptId().getApplicationId()

          + " launched container " + containerId + " on node: " + node);

      return;

    }

    application.containerLaunchedOnNode(containerId, node.getNodeID());

  }

  /**

   * Process a heartbeat update from a node.

   */

  private synchronized void nodeUpdate(RMNode nm) {

    if (LOG.isDebugEnabled()) {

      LOG.debug("nodeUpdate: " + nm + " cluster capacity: " + clusterCapacity);

    }

    eventLog.log("HEARTBEAT", nm.getHostName());

    FSSchedulerNode node = nodes.get(nm.getNodeID());

    // Update resource if any change

    SchedulerUtils.updateResourceIfChanged(node, nm, clusterCapacity, LOG);

   

    List<UpdatedContainerInfo> containerInfoList = nm.pullContainerUpdates();

    List<ContainerStatus> newlyLaunchedContainers = new ArrayList<ContainerStatus>();

    List<ContainerStatus> completedContainers = new ArrayList<ContainerStatus>();

    for(UpdatedContainerInfo containerInfo : containerInfoList) {

      newlyLaunchedContainers.addAll(containerInfo.getNewlyLaunchedContainers());

      completedContainers.addAll(containerInfo.getCompletedContainers());

    }

    // Processing the newly launched containers

    for (ContainerStatus launchedContainer : newlyLaunchedContainers) {

      containerLaunchedOnNode(launchedContainer.getContainerId(), node);

    }

    // Process completed containers

    for (ContainerStatus completedContainer : completedContainers) {

      ContainerId containerId = completedContainer.getContainerId();

      LOG.debug("Container FINISHED: " + containerId);

      completedContainer(getRMContainer(containerId),

          completedContainer, RMContainerEventType.FINISHED);

    }

    if (continuousSchedulingEnabled) {

      if (!completedContainers.isEmpty()) {

        attemptScheduling(node);

      }

    } else {

      attemptScheduling(node);

    }

  }

  private void continuousScheduling() {

    while (true) {

      List<NodeId> nodeIdList = new ArrayList<NodeId>(nodes.keySet());

      // Sort the nodes by space available on them, so that we offer

      // containers on emptier nodes first, facilitating an even spread. This

      // requires holding the scheduler lock, so that the space available on a

      // node doesn't change during the sort.

      synchronized (this) {

        Collections.sort(nodeIdList, nodeAvailableResourceComparator);

      }

      // iterate all nodes

      for (NodeId nodeId : nodeIdList) {

        if (nodes.containsKey(nodeId)) {

          FSSchedulerNode node = nodes.get(nodeId);

          try {

            if (Resources.fitsIn(minimumAllocation,

                    node.getAvailableResource())) {

              attemptScheduling(node);

            }

          } catch (Throwable ex) {

            LOG.warn("Error while attempting scheduling for node " + node +

                    ": " + ex.toString(), ex);

          }

        }

      }

      try {

        Thread.sleep(getContinuousSchedulingSleepMs());

      } catch (InterruptedException e) {

        LOG.warn("Error while doing sleep in continuous scheduling: " +

                e.toString(), e);

      }

    }

  }

  /** Sort nodes by available resource */

  private class NodeAvailableResourceComparator implements Comparator<NodeId> {

    @Override

    public int compare(NodeId n1, NodeId n2) {

      return RESOURCE_CALCULATOR.compare(clusterCapacity,

              nodes.get(n2).getAvailableResource(),

              nodes.get(n1).getAvailableResource());

    }

  }

 

  private synchronized void attemptScheduling(FSSchedulerNode node) {

    // Assign new containers...

    // 1. Check for reserved applications

    // 2. Schedule if there are no reservations

    AppSchedulable reservedAppSchedulable = node.getReservedAppSchedulable();

    if (reservedAppSchedulable != null) {

      Priority reservedPriority = node.getReservedContainer().getReservedPriority();

      if (!reservedAppSchedulable.hasContainerForNode(reservedPriority, node)) {

        // Don't hold the reservation if app can no longer use it

        LOG.info("Releasing reservation that cannot be satisfied for application "

            + reservedAppSchedulable.getApp().getApplicationAttemptId()

            + " on node " + node);

        reservedAppSchedulable.unreserve(reservedPriority, node);

        reservedAppSchedulable = null;

      } else {

        // Reservation exists; try to fulfill the reservation

        LOG.info("Trying to fulfill reservation for application "

            + reservedAppSchedulable.getApp().getApplicationAttemptId()

            + " on node: " + node);

        node.getReservedAppSchedulable().assignReservedContainer(node);

      }

    }

    if (reservedAppSchedulable == null) {

      // No reservation, schedule at queue which is farthest below fair share

      int assignedContainers = 0;

      while (node.getReservedContainer() == null) {

        boolean assignedContainer = false;

        if (Resources.greaterThan(RESOURCE_CALCULATOR, clusterCapacity,

              queueMgr.getRootQueue().assignContainer(node),

              Resources.none())) {

          assignedContainers++;

          assignedContainer = true;

        }

        if (!assignedContainer) { break; }

        if (!assignMultiple) { break; }

        if ((assignedContainers >= maxAssign) && (maxAssign > 0)) { break; }

      }

    }

    updateRootQueueMetrics();

  }

  @Override

  public SchedulerNodeReport getNodeReport(NodeId nodeId) {

    FSSchedulerNode node = nodes.get(nodeId);

    return node == null ? null : new SchedulerNodeReport(node);

  }

 

  public FSSchedulerApp getSchedulerApp(ApplicationAttemptId appAttemptId) {

    SchedulerApplication app =

        applications.get(appAttemptId.getApplicationId());

    if (app != null) {

      return (FSSchedulerApp) app.getCurrentAppAttempt();

    }

    return null;

  }

 

  @Override

  public SchedulerAppReport getSchedulerAppInfo(

      ApplicationAttemptId appAttemptId) {

    FSSchedulerApp attempt = getSchedulerApp(appAttemptId);

    if (attempt == null) {

      if (LOG.isDebugEnabled()) {

        LOG.debug("Request for appInfo of unknown attempt " + appAttemptId);

      }

      return null;

    }

    return new SchedulerAppReport(attempt);

  }

 

  @Override

  public ApplicationResourceUsageReport getAppResourceUsageReport(

      ApplicationAttemptId appAttemptId) {

    FSSchedulerApp attempt = getSchedulerApp(appAttemptId);

    if (attempt == null) {

      if (LOG.isDebugEnabled()) {

        LOG.debug("Request for appInfo of unknown attempt " + appAttemptId);

      }

      return null;

    }

    return attempt.getResourceUsageReport();

  }

 

  /**

   * Subqueue metrics might be a little out of date because fair shares are

   * recalculated at the update interval, but the root queue metrics needs to

   * be updated synchronously with allocations and completions so that cluster

   * metrics will be consistent.

   */

  private void updateRootQueueMetrics() {

    rootMetrics.setAvailableResourcesToQueue(

        Resources.subtract(

            clusterCapacity, rootMetrics.getAllocatedResources()));

  }

  @Override

  public QueueMetrics getRootQueueMetrics() {

    return rootMetrics;

  }

  @Override

  public void handle(SchedulerEvent event) {

    switch (event.getType()) {

    case NODE_ADDED:

      if (!(event instanceof NodeAddedSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      NodeAddedSchedulerEvent nodeAddedEvent = (NodeAddedSchedulerEvent)event;

      addNode(nodeAddedEvent.getAddedRMNode());

      break;

    case NODE_REMOVED:

      if (!(event instanceof NodeRemovedSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      NodeRemovedSchedulerEvent nodeRemovedEvent = (NodeRemovedSchedulerEvent)event;

      removeNode(nodeRemovedEvent.getRemovedRMNode());

      break;

    case NODE_UPDATE:

      if (!(event instanceof NodeUpdateSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      NodeUpdateSchedulerEvent nodeUpdatedEvent = (NodeUpdateSchedulerEvent)event;

      nodeUpdate(nodeUpdatedEvent.getRMNode());

      break;

    case APP_ADDED:

      if (!(event instanceof AppAddedSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      AppAddedSchedulerEvent appAddedEvent = (AppAddedSchedulerEvent) event;

      addApplication(appAddedEvent.getApplicationId(),

        appAddedEvent.getQueue(), appAddedEvent.getUser());

      break;

    case APP_REMOVED:

      if (!(event instanceof AppRemovedSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      AppRemovedSchedulerEvent appRemovedEvent = (AppRemovedSchedulerEvent)event;

      removeApplication(appRemovedEvent.getApplicationID(),

        appRemovedEvent.getFinalState());

      break;

    case APP_ATTEMPT_ADDED:

      if (!(event instanceof AppAttemptAddedSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      AppAttemptAddedSchedulerEvent appAttemptAddedEvent =

          (AppAttemptAddedSchedulerEvent) event;

      addApplicationAttempt(appAttemptAddedEvent.getApplicationAttemptId(),

        appAttemptAddedEvent.getTransferStateFromPreviousAttempt());

      break;

    case APP_ATTEMPT_REMOVED:

      if (!(event instanceof AppAttemptRemovedSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      AppAttemptRemovedSchedulerEvent appAttemptRemovedEvent =

          (AppAttemptRemovedSchedulerEvent) event;

      removeApplicationAttempt(

          appAttemptRemovedEvent.getApplicationAttemptID(),

          appAttemptRemovedEvent.getFinalAttemptState(),

          appAttemptRemovedEvent.getKeepContainersAcrossAppAttempts());

      break;

    case CONTAINER_EXPIRED:

      if (!(event instanceof ContainerExpiredSchedulerEvent)) {

        throw new RuntimeException("Unexpected event type: " + event);

      }

      ContainerExpiredSchedulerEvent containerExpiredEvent =

          (ContainerExpiredSchedulerEvent)event;

      ContainerId containerId = containerExpiredEvent.getContainerId();

      completedContainer(getRMContainer(containerId),

          SchedulerUtils.createAbnormalContainerStatus(

              containerId,

              SchedulerUtils.EXPIRED_CONTAINER),

          RMContainerEventType.EXPIRE);

      break;

    default:

      LOG.error("Unknown event arrived at FairScheduler: " + event.toString());

    }

  }

  @Override

  public void recover(RMState state) throws Exception {

    // NOT IMPLEMENTED

  }

  @Override

  public synchronized void reinitialize(Configuration conf, RMContext rmContext)

      throws IOException {

    if (!initialized) {

      this.conf = new FairSchedulerConfiguration(conf);

      validateConf(this.conf);

      minimumAllocation = this.conf.getMinimumAllocation();

      maximumAllocation = this.conf.getMaximumAllocation();

      incrAllocation = this.conf.getIncrementAllocation();

      continuousSchedulingEnabled = this.conf.isContinuousSchedulingEnabled();

      continuousSchedulingSleepMs =

              this.conf.getContinuousSchedulingSleepMs();

      nodeLocalityThreshold = this.conf.getLocalityThresholdNode();

      rackLocalityThreshold = this.conf.getLocalityThresholdRack();

      nodeLocalityDelayMs = this.conf.getLocalityDelayNodeMs();

      rackLocalityDelayMs = this.conf.getLocalityDelayRackMs();

      preemptionEnabled = this.conf.getPreemptionEnabled();

      assignMultiple = this.conf.getAssignMultiple();

      maxAssign = this.conf.getMaxAssign();

      sizeBasedWeight = this.conf.getSizeBasedWeight();

      preemptionInterval = this.conf.getPreemptionInterval();

      waitTimeBeforeKill = this.conf.getWaitTimeBeforeKill();

      usePortForNodeName = this.conf.getUsePortForNodeName();

     

      rootMetrics = FSQueueMetrics.forQueue("root", null, true, conf);

      this.rmContext = rmContext;

      // This stores per-application scheduling information

      this.applications =

          new ConcurrentHashMap<ApplicationId, SchedulerApplication>();

      this.eventLog = new FairSchedulerEventLog();

      eventLog.init(this.conf);

      initialized = true;

      allocConf = new AllocationConfiguration(conf);

      try {

        queueMgr.initialize(conf);

      } catch (Exception e) {

        throw new IOException("Failed to start FairScheduler", e);

      }

      Thread updateThread = new Thread(new UpdateThread());

      updateThread.setName("FairSchedulerUpdateThread");

      updateThread.setDaemon(true);

      updateThread.start();

      if (continuousSchedulingEnabled) {

        // start continuous scheduling thread

        Thread schedulingThread = new Thread(

          new Runnable() {

            @Override

            public void run() {

              continuousScheduling();

            }

          }

        );

        schedulingThread.setName("ContinuousScheduling");

        schedulingThread.setDaemon(true);

        schedulingThread.start();

      }

     

      allocsLoader.init(conf);

      allocsLoader.setReloadListener(new AllocationReloadListener());

      // If we fail to load allocations file on initialize, we want to fail

      // immediately.  After a successful load, exceptions on future reloads

      // will just result in leaving things as they are.

      try {

        allocsLoader.reloadAllocations();

      } catch (Exception e) {

        throw new IOException("Failed to initialize FairScheduler", e);

      }

      allocsLoader.start();

    } else {

      try {

        allocsLoader.reloadAllocations();

      } catch (Exception e) {

        LOG.error("Failed to reload allocations file", e);

      }

    }

  }

  @Override

  public QueueInfo getQueueInfo(String queueName, boolean includeChildQueues,

      boolean recursive) throws IOException {

    if (!queueMgr.exists(queueName)) {

      throw new IOException("queue " + queueName + " does not exist");

    }

    return queueMgr.getQueue(queueName).getQueueInfo(includeChildQueues,

        recursive);

  }

  @Override

  public List<QueueUserACLInfo> getQueueUserAclInfo() {

    UserGroupInformation user = null;

    try {

      user = UserGroupInformation.getCurrentUser();

    } catch (IOException ioe) {

      return new ArrayList<QueueUserACLInfo>();

    }

    return queueMgr.getRootQueue().getQueueUserAclInfo(user);

  }

  @Override

  public int getNumClusterNodes() {

    return nodes.size();

  }

  @Override

  public synchronized boolean checkAccess(UserGroupInformation callerUGI,

      QueueACL acl, String queueName) {

    FSQueue queue = getQueueManager().getQueue(queueName);

    if (queue == null) {

      if (LOG.isDebugEnabled()) {

        LOG.debug("ACL not found for queue access-type " + acl

            + " for queue " + queueName);

      }

      return false;

    }

    return queue.hasAccess(acl, callerUGI);

  }

 

  public AllocationConfiguration getAllocationConfiguration() {

    return allocConf;

  }

 

  private class AllocationReloadListener implements

      AllocationFileLoaderService.Listener {

    @Override

    public void onReload(AllocationConfiguration queueInfo) {

      // Commit the reload; also create any queue defined in the alloc file

      // if it does not already exist, so it can be displayed on the web UI.

      synchronized (FairScheduler.this) {

        allocConf = queueInfo;

        allocConf.getDefaultSchedulingPolicy().initialize(clusterCapacity);

        queueMgr.updateAllocationConfiguration(allocConf);

      }

    }

  }

  @Override

  public List<ApplicationAttemptId> getAppsInQueue(String queueName) {

    FSQueue queue = queueMgr.getQueue(queueName);

    if (queue == null) {

      return null;

    }

    List<ApplicationAttemptId> apps = new ArrayList<ApplicationAttemptId>();

    queue.collectSchedulerApplications(apps);

    return apps;

  }

  @Override

  public synchronized String moveApplication(ApplicationId appId,

      String queueName) throws YarnException {

    SchedulerApplication app = applications.get(appId);