YARN-10903. Fix the headroom check in ParentQueue and RegularContainerAllocator for DRF (#3352)

Contributed by Jie Wang <jie.wang@hulu.com>

Author: JackWangCS

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/ParentQueue.java

@@ -1038,10 +1038,9 @@ private boolean canAssign(Resource clusterResource, FiCaSchedulerNode node) {
     // Two conditions need to meet when trying to allocate:
     // 1) Node doesn't have reserved container
     // 2) Node's available-resource + killable-resource should > 0
-    boolean accept = node.getReservedContainer() == null && Resources
-        .greaterThanOrEqual(resourceCalculator, clusterResource, Resources
-            .add(node.getUnallocatedResource(),
-                node.getTotalKillableResources()), minimumAllocation);
+    boolean accept = node.getReservedContainer() == null &&
+        Resources.fitsIn(resourceCalculator, minimumAllocation,
+            Resources.add(node.getUnallocatedResource(), node.getTotalKillableResources()));
     if (!accept) {
       ActivitiesLogger.QUEUE.recordQueueActivity(activitiesManager, node,
           getParentName(), getQueuePath(), ActivityState.REJECTED,

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/allocator/RegularContainerAllocator.java

@@ -73,10 +73,9 @@ public RegularContainerAllocator(FiCaSchedulerApp application,
       ActivitiesManager activitiesManager) {
     super(application, rc, rmContext, activitiesManager);
   }
-  
-  private boolean checkHeadroom(Resource clusterResource,
-      ResourceLimits currentResourceLimits, Resource required,
-      String nodePartition) {
+
+  private boolean checkHeadroom(ResourceLimits currentResourceLimits,
+                                Resource required, String nodePartition) {
     // If headroom + currentReservation < required, we cannot allocate this
     // require
     Resource resourceCouldBeUnReserved =
@@ -86,9 +85,8 @@ private boolean checkHeadroom(Resource clusterResource,
       // we won't allow to unreserve before allocation.
       resourceCouldBeUnReserved = Resources.none();
     }
-    return Resources.greaterThanOrEqual(rc, clusterResource, Resources.add(
-        currentResourceLimits.getHeadroom(), resourceCouldBeUnReserved),
-        required);
+    return Resources.fitsIn(rc, required,
+        Resources.add(currentResourceLimits.getHeadroom(), resourceCouldBeUnReserved));
   }
 
   /*
@@ -97,8 +95,7 @@ private boolean checkHeadroom(Resource clusterResource,
    * We will consider stuffs like exclusivity, pending resource, node partition,
    * headroom, etc.
    */
-  private ContainerAllocation preCheckForNodeCandidateSet(
-      Resource clusterResource, FiCaSchedulerNode node,
+  private ContainerAllocation preCheckForNodeCandidateSet(FiCaSchedulerNode node,
       SchedulingMode schedulingMode, ResourceLimits resourceLimits,
       SchedulerRequestKey schedulerKey) {
     PendingAsk offswitchPendingAsk = application.getPendingAsk(schedulerKey,
@@ -168,8 +165,7 @@ private ContainerAllocation preCheckForNodeCandidateSet(
       }
     }
 
-    if (!checkHeadroom(clusterResource, resourceLimits, required,
-        node.getPartition())) {
+    if (!checkHeadroom(resourceLimits, required, node.getPartition())) {
       LOG.debug("cannot allocate required resource={} because of headroom",
           required);
       ActivitiesLogger.APP.recordAppActivityWithoutAllocation(
@@ -857,7 +853,7 @@ private ContainerAllocation allocate(Resource clusterResource,
       FiCaSchedulerNode node = iter.next();
 
       if (reservedContainer == null) {
-        result = preCheckForNodeCandidateSet(clusterResource, node,
+        result = preCheckForNodeCandidateSet(node,
             schedulingMode, resourceLimits, schedulerKey);
         if (null != result) {
           continue;

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/TestLeafQueue.java

@@ -800,6 +800,62 @@ public void testSingleQueueWithOneUser() throws Exception {
     assertEquals((int)(a.getCapacity() * node_0.getTotalResource().getMemorySize()),
         a.getMetrics().getAvailableMB());
   }
+
+  @Test
+  public void testHeadroomCheckWithDRF() throws Exception {
+    CSAssignment assignment;
+    setUpWithDominantResourceCalculator();
+    // Mock the queue
+    LeafQueue b = stubLeafQueue((LeafQueue) queues.get(B));
+    // Users
+    final String user0 = "user_0";
+
+    // Submit applications
+    final ApplicationAttemptId appAttemptId0 =
+        TestUtils.getMockApplicationAttemptId(0, 0);
+    FiCaSchedulerApp app0 =
+        new FiCaSchedulerApp(appAttemptId0, user0, b,
+            b.getAbstractUsersManager(), spyRMContext);
+    b.submitApplicationAttempt(app0, user0);
+    // Setup some nodes
+    String host0 = "127.0.0.1";
+    FiCaSchedulerNode node0 =
+        TestUtils.getMockNode(host0, DEFAULT_RACK, 0, 100 * GB, 100);
+
+    int numNodes = 1;
+    Resource clusterResource =
+        Resources.createResource(numNodes * (100 * GB), numNodes * 100);
+    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
+    root.updateClusterResource(clusterResource,
+        new ResourceLimits(clusterResource));
+
+    // Increase the user-limit-factor to make user_0 fully use max resources of the queue.
+    // The max resources can be used are 0.99 * [100 * GB, 100]
+    b.setUserLimitFactor(10.0f);
+
+    Map<ApplicationAttemptId, FiCaSchedulerApp> apps =
+        ImmutableMap.of(app0.getApplicationAttemptId(), app0);
+    Map<NodeId, FiCaSchedulerNode> nodes = ImmutableMap.of(node0.getNodeID(), node0);
+
+    Priority priority = TestUtils.createMockPriority(1);
+    app0.updateResourceRequests(Collections.singletonList(TestUtils
+        .createResourceRequest(ResourceRequest.ANY, 90 * GB, 10, 1, true,
+            priority, recordFactory, NO_LABEL)));
+    assignment = b.assignContainers(clusterResource, node0, new ResourceLimits(
+        clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+    applyCSAssignment(clusterResource, assignment, b, nodes, apps);
+    verifyContainerAllocated(assignment, NodeType.OFF_SWITCH);
+
+    app0.updateResourceRequests(Collections.singletonList(TestUtils
+        .createResourceRequest(ResourceRequest.ANY, 10 * GB, 10, 1, true,
+            priority, recordFactory, NO_LABEL)));
+    assignment = b.assignContainers(clusterResource, node0, new ResourceLimits(
+        clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+    // This assignment should have no containers assigned,
+    // because the used memory (90 + 10)GB will exceed the max 99GB
+    verifyNoContainerAllocated(assignment);
+  }
+
   @Test
   public void testDRFUsageRatioRounding() throws Exception {
     CSAssignment assign;