<Yarn> <Capacity Scheduler> <Source Code>

Yarn capacity scheduler

• 首先要知道， [Attention: RM有两个组件，其中Scheduler完全就只是负责资源的分配；ApplicationsManager则负责接受application，选取ApplicationMaster，监控重启AM。]
• CapacityScheduler的优点就是灵活，集群的利用率高；缺点也是由其灵活性造成的，因为CapacityScheduler不支持抢占式调度，必须等上一个任务主动释放资源。
• 
• 可以看出，只要提交的app数没有达到上限，就能够一直提交app到集群，只是这些app会处于accepted的状态，一直pending，直到ResourceManager给它分配资源。
• 因此Scheduler实际上是会不断地schedule，然后分配资源给那些集群，在具体分配的时候，在queue内部可能会考虑优先级，然后结合FIFO。

AsyncScheduleThread

• 在CapacityScheduler内部有个AsyncScheduleThread 这个异步Scheduler线程会不断地调schedule(cs)方法。

schedule(cs)

• // Schedule on all nodes by starting at a random point.
• static void schedule(CapacityScheduler cs) 
  • // first randomize the start point
  • Collection<FiCaSchedulerNode> nodes = cs.getAllNodes().values()    // get all the nodes in the cluster
  • for each node: cs.allocateContainersToNode(node)
    • // assign new containers...  1. check for reserved apps(for reservation see here and here. )   2. schedule if there are no reservations
    • if reservedContainer != null
      • get the reserved apps according to reservedContainers
      • // try to fulfill the reservation
      • LeafQueue queue = ((LeafQueue) reservedApplication.getQueue();
      • CSAssignment assignment = queue.assignContainers(clusterResource, node, false); // assignContainers(Resource clusterResource, FiCaSchedulerNode node, boolean needToUnreserve);
        • // if our queue cannot access this node, just return
        • // check for reserved resources
      • TBD...
    • // try to schedule more if there are no reservations to fulfill
    • if (node.getReservedContainer() == null)
      • if (calculator.computeAvailableContainers(node.getAvailableResource(), minimumAllocation) > 0), then  // computeAvailableContainers(Resource available, Resource required), as for DominantResourceCalculator, return the min ratio of mem & vcores.
        • assignContainers(clusterResource, node, false)   // assignContainers(Resource clusterResource, FiCaSchedulerNode node)
          • // if our queue cannot access this node, just return
          • // check for reserved resources    // TBD...
          • // try to assign containers to apps in order
          • for (FoCaSchedulerApp application: activeApplications) 
            • if (SchedulerAppUtils.isBlacklisted(application, node, LOG) // check if this resource is on blacklist, i.e. cannot run the app in this node/rack
            • // schedule in priority order, this is the priority of the resourceRequest of this app
            • for (Priority priority : application.getPriorities())
              • ResourceRequest anyRequest = application.getResourceRequest(priority, ResourceRequest.ANY)
              • Resource required = anyRequest.getCapacity()
              • Set<String> requestedNodeLabels = getRequestLabelSetByExpression(anyRequest.getNodeLabelExpression());
              • // compute user-limit & set headroom. 
              • Resource userLimit = computeUserLimitAndSetHeadroom(application, clusterResource, required, requestedNodeLabels)
                • // compute user limit respect requested labels
                • // TODO: need consider headroom respect labels also
                • Resource userLimit = computeUserLimit(application, clusterResource, required, queueUser, requestedLabels)
                  • // our current capacity: equal to the max(required, queue-capacity) if we're running below capacity, equal to (usedResources + required) if running over capacity.
                  • // if we have labels to request(choose to use the first one).
                  • // else if no label on request, just use absolute capacity as capacity for nodes without label.
                  • // TBD...
              • // max avail capacity needs to take into account usage by ancestor-siblings which are greater than their base 
              • // calculate absoluteMaxAvailCapacity: my max avail is min(my max capacity, unused from my parent by my siblings if they are beyond their base capacity)
              • // then calculate queueMaxCap using absoluteMaxAvailCapacity
              • // check canAssignToThisQueue
                • // consider the intersection of queue-canAccessLabels and node-labels, if any of the label beyond queue limit, we cannot allocate on this node.
                • // check user limit
                • application.addSchedulingOpportunity(priority);
                • // try to schedule...
                • TBD...
• FYI:
  

  /** 
   * Headroom is:
   *    min(
   *        min(userLimit, queueMaxCap) - userConsumed,
   *        queueMaxCap - queueUsedResources
   *       )
   * 
   * ( which can be expressed as, 
   *  min (userLimit - userConsumed, queuMaxCap - userConsumed, 
   *    queueMaxCap - queueUsedResources)
   *  )
   *
   * given that queueUsedResources >= userConsumed, this simplifies to
   *
   * >> min (userlimit - userConsumed,   queueMaxCap - queueUsedResources) << 
   *
   */

addApplication

首先在CapacityScheduler随意找了个方法

synchronized addApplication(ApplicationAttemptedId applicationAttemptId, String queueName, String user)

• sanity check
  • queue == null
  • !queue instanceof LeafQueue
• Represents an application from the viewpoint of the scheduler. (Each running app in the RM corresponds to one instance of the FiCaScheduler class)
• FiCaSchedulerApp SchedulerApp = new FiCaSchedulerApp(applicationAttemptId, user, queue, queue.getActiveUserManager(), rmContext);
  • ActiveUsersManager tracks users in the system. (An active user is defined as someone with outstanding resource requests.)
  • rmContext is the context of the RM.
• submit to the queue
  • try: queue.submitApplication(SchedulerApp, user, queueName)
    • check queue ACLs
    • synchronized(this) 
      • check if the queue is accepting jobs: if (getState() != QueueState.RUNNING) throw Exception
      • check submission limits for queues:
        • if (getNumApplications() >= getMaxApplications()) throw Exception
      • check submission limits for the user on this queue
      • addApplication(applictaion, user)
        • user.submitApplication() : pendingApp ++; // accepted
        • activateApplications():
          • for each pending apps: 
            • check queue limit & user limit again (same as above)
            • activateApplication(): --pendingApp; ++activeApp;
    • metrics.submitApp(userName, attempId): // each queue has a metrics which is an instance of QueueMetrics
      • update metrics: appsSubmitted, appsFailed, appsPending
      • if (parent != null) parent.submitApp(user, attemptId)   // to inform the parents recursively

以上，可以看到在addApplication方法内主要是判断了ACL和appNum的上限，没有resource相关的分配和判断。资源（container）的分配是由相应的applicationMaster向Resourcemanager统一请求的。ResourceRequest使用protobuf。

• 用户提交应用程序 --> ResourceManager   --> ACL等检查 -->  app accepted.
• 一旦Scheduler有足够的资源可以满足需求  -->  app由accepted转成running  -->  RM为ApplicationMaster分配一个container，并负责在节点上拉起它。
• AM是每个用户作业的主进程，负责管理作业生命周期，包括动态地增加or减少资源(container)，管理执行流程，处理故障和计算偏差。 

Yarn Queues