Examples with FiCaSchedulerNode org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode used on opensource projects

Example 71 with FiCaSchedulerNode

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode in project hadoop by apache.

the class TestReservations method testReservationNoContinueLook.

@Test
public void testReservationNoContinueLook() throws Exception {
    // Test that with reservations-continue-look-all-nodes feature off
    // we don't unreserve and show we could get stuck
    queues = new HashMap<String, CSQueue>();
    // test that the deadlock occurs when turned off
    CapacitySchedulerConfiguration csConf = new CapacitySchedulerConfiguration();
    csConf.setBoolean(CapacitySchedulerConfiguration.RESERVE_CONT_LOOK_ALL_NODES, false);
    setup(csConf);
    // Manipulate queue 'a'
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(A));
    // Users
    final String user_0 = "user_0";
    // Submit applications
    final ApplicationAttemptId appAttemptId_0 = TestUtils.getMockApplicationAttemptId(0, 0);
    FiCaSchedulerApp app_0 = new FiCaSchedulerApp(appAttemptId_0, user_0, a, mock(ActiveUsersManager.class), spyRMContext);
    app_0 = spy(app_0);
    Mockito.doNothing().when(app_0).updateAMContainerDiagnostics(any(AMState.class), any(String.class));
    rmContext.getRMApps().put(app_0.getApplicationId(), mock(RMApp.class));
    a.submitApplicationAttempt(app_0, user_0);
    final ApplicationAttemptId appAttemptId_1 = TestUtils.getMockApplicationAttemptId(1, 0);
    FiCaSchedulerApp app_1 = new FiCaSchedulerApp(appAttemptId_1, user_0, a, mock(ActiveUsersManager.class), spyRMContext);
    app_1 = spy(app_1);
    Mockito.doNothing().when(app_1).updateAMContainerDiagnostics(any(AMState.class), any(String.class));
    a.submitApplicationAttempt(app_1, user_0);
    // Setup some nodes
    String host_0 = "host_0";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, DEFAULT_RACK, 0, 8 * GB);
    String host_1 = "host_1";
    FiCaSchedulerNode node_1 = TestUtils.getMockNode(host_1, DEFAULT_RACK, 0, 8 * GB);
    String host_2 = "host_2";
    FiCaSchedulerNode node_2 = TestUtils.getMockNode(host_2, DEFAULT_RACK, 0, 8 * GB);
    Map<ApplicationAttemptId, FiCaSchedulerApp> apps = ImmutableMap.of(app_0.getApplicationAttemptId(), app_0, app_1.getApplicationAttemptId(), app_1);
    Map<NodeId, FiCaSchedulerNode> nodes = ImmutableMap.of(node_0.getNodeID(), node_0, node_1.getNodeID(), node_1, node_2.getNodeID(), node_2);
    when(csContext.getNode(node_0.getNodeID())).thenReturn(node_0);
    when(csContext.getNode(node_1.getNodeID())).thenReturn(node_1);
    when(csContext.getNode(node_2.getNodeID())).thenReturn(node_2);
    final int numNodes = 3;
    Resource clusterResource = Resources.createResource(numNodes * (8 * GB));
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Setup resource-requests
    Priority priorityAM = TestUtils.createMockPriority(1);
    Priority priorityMap = TestUtils.createMockPriority(5);
    Priority priorityReduce = TestUtils.createMockPriority(10);
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 2 * GB, 1, true, priorityAM, recordFactory)));
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 5 * GB, 2, true, priorityReduce, recordFactory)));
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 3 * GB, 2, true, priorityMap, recordFactory)));
    // Start testing...
    // Only AM
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(2 * GB, a.getUsedResources().getMemorySize());
    assertEquals(2 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, a.getMetrics().getReservedMB());
    assertEquals(2 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(22 * GB, a.getMetrics().getAvailableMB());
    assertEquals(2 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_2.getAllocatedResource().getMemorySize());
    // Only 1 map - simulating reduce
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(5 * GB, a.getUsedResources().getMemorySize());
    assertEquals(5 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, a.getMetrics().getReservedMB());
    assertEquals(5 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(19 * GB, a.getMetrics().getAvailableMB());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_2.getAllocatedResource().getMemorySize());
    // Only 1 map to other node - simulating reduce
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(8 * GB, a.getUsedResources().getMemorySize());
    assertEquals(8 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, a.getMetrics().getReservedMB());
    assertEquals(8 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(16 * GB, a.getMetrics().getAvailableMB());
    assertEquals(16 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(null, node_0.getReservedContainer());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(3 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_2.getAllocatedResource().getMemorySize());
    assertEquals(2, app_0.getOutstandingAsksCount(toSchedulerKey(priorityReduce)));
    // try to assign reducer (5G on node 0 and should reserve)
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(13 * GB, a.getUsedResources().getMemorySize());
    assertEquals(8 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(5 * GB, a.getMetrics().getReservedMB());
    assertEquals(8 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(11 * GB, a.getMetrics().getAvailableMB());
    assertEquals(11 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(5 * GB, node_0.getReservedContainer().getReservedResource().getMemorySize());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(3 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_2.getAllocatedResource().getMemorySize());
    assertEquals(2, app_0.getOutstandingAsksCount(toSchedulerKey(priorityReduce)));
    // assign reducer to node 2
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(18 * GB, a.getUsedResources().getMemorySize());
    assertEquals(13 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(5 * GB, a.getMetrics().getReservedMB());
    assertEquals(13 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(6 * GB, a.getMetrics().getAvailableMB());
    assertEquals(6 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(5 * GB, node_0.getReservedContainer().getReservedResource().getMemorySize());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(3 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(5 * GB, node_2.getAllocatedResource().getMemorySize());
    assertEquals(1, app_0.getOutstandingAsksCount(toSchedulerKey(priorityReduce)));
    // node_1 heartbeat and won't unreserve from node_0, potentially stuck
    // if AM doesn't handle
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(18 * GB, a.getUsedResources().getMemorySize());
    assertEquals(13 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(5 * GB, a.getMetrics().getReservedMB());
    assertEquals(13 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(6 * GB, a.getMetrics().getAvailableMB());
    assertEquals(6 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(5 * GB, node_0.getReservedContainer().getReservedResource().getMemorySize());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(3 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(5 * GB, node_2.getAllocatedResource().getMemorySize());
    assertEquals(1, app_0.getOutstandingAsksCount(toSchedulerKey(priorityReduce)));
}

Example 72 with FiCaSchedulerNode

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode in project hadoop by apache.

the class TestReservations method testGetAppToUnreserve.

@Test
public void testGetAppToUnreserve() throws Exception {
    CapacitySchedulerConfiguration csConf = new CapacitySchedulerConfiguration();
    setup(csConf);
    final String user_0 = "user_0";
    final ApplicationAttemptId appAttemptId_0 = TestUtils.getMockApplicationAttemptId(0, 0);
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(A));
    FiCaSchedulerApp app_0 = new FiCaSchedulerApp(appAttemptId_0, user_0, a, mock(ActiveUsersManager.class), spyRMContext);
    String host_0 = "host_0";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, DEFAULT_RACK, 0, 8 * GB);
    String host_1 = "host_1";
    FiCaSchedulerNode node_1 = TestUtils.getMockNode(host_1, DEFAULT_RACK, 0, 8 * GB);
    Resource clusterResource = Resources.createResource(2 * 8 * GB);
    // Setup resource-requests
    Priority p = TestUtils.createMockPriority(5);
    SchedulerRequestKey priorityMap = toSchedulerKey(p);
    Resource capability = Resources.createResource(2 * GB, 0);
    RMApplicationHistoryWriter writer = mock(RMApplicationHistoryWriter.class);
    SystemMetricsPublisher publisher = mock(SystemMetricsPublisher.class);
    RMContext rmContext = mock(RMContext.class);
    ContainerAllocationExpirer expirer = mock(ContainerAllocationExpirer.class);
    DrainDispatcher drainDispatcher = new DrainDispatcher();
    when(rmContext.getContainerAllocationExpirer()).thenReturn(expirer);
    when(rmContext.getDispatcher()).thenReturn(drainDispatcher);
    when(rmContext.getRMApplicationHistoryWriter()).thenReturn(writer);
    when(rmContext.getSystemMetricsPublisher()).thenReturn(publisher);
    when(rmContext.getYarnConfiguration()).thenReturn(new YarnConfiguration());
    ApplicationAttemptId appAttemptId = BuilderUtils.newApplicationAttemptId(app_0.getApplicationId(), 1);
    ContainerId containerId = BuilderUtils.newContainerId(appAttemptId, 1);
    Container container = TestUtils.getMockContainer(containerId, node_1.getNodeID(), Resources.createResource(2 * GB), priorityMap.getPriority());
    RMContainer rmContainer = new RMContainerImpl(container, SchedulerRequestKey.extractFrom(container), appAttemptId, node_1.getNodeID(), "user", rmContext);
    Container container_1 = TestUtils.getMockContainer(containerId, node_0.getNodeID(), Resources.createResource(1 * GB), priorityMap.getPriority());
    RMContainer rmContainer_1 = new RMContainerImpl(container_1, SchedulerRequestKey.extractFrom(container_1), appAttemptId, node_0.getNodeID(), "user", rmContext);
    // no reserved containers
    NodeId unreserveId = app_0.getNodeIdToUnreserve(priorityMap, capability, cs.getResourceCalculator(), clusterResource);
    assertEquals(null, unreserveId);
    // no reserved containers - reserve then unreserve
    app_0.reserve(node_0, priorityMap, rmContainer_1, container_1);
    app_0.unreserve(priorityMap, node_0, rmContainer_1);
    unreserveId = app_0.getNodeIdToUnreserve(priorityMap, capability, cs.getResourceCalculator(), clusterResource);
    assertEquals(null, unreserveId);
    // no container large enough is reserved
    app_0.reserve(node_0, priorityMap, rmContainer_1, container_1);
    unreserveId = app_0.getNodeIdToUnreserve(priorityMap, capability, cs.getResourceCalculator(), clusterResource);
    assertEquals(null, unreserveId);
    // reserve one that is now large enough
    app_0.reserve(node_1, priorityMap, rmContainer, container);
    unreserveId = app_0.getNodeIdToUnreserve(priorityMap, capability, cs.getResourceCalculator(), clusterResource);
    assertEquals(node_1.getNodeID(), unreserveId);
}

Example 73 with FiCaSchedulerNode

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode in project hadoop by apache.

the class TestReservations method testAssignContainersNeedToUnreserve.

@Test
@SuppressWarnings("unchecked")
public void testAssignContainersNeedToUnreserve() throws Exception {
    // Test that we now unreserve and use a node that has space
    Logger rootLogger = LogManager.getRootLogger();
    rootLogger.setLevel(Level.DEBUG);
    CapacitySchedulerConfiguration csConf = new CapacitySchedulerConfiguration();
    setup(csConf);
    // Manipulate queue 'a'
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(A));
    // Users
    final String user_0 = "user_0";
    // Submit applications
    final ApplicationAttemptId appAttemptId_0 = TestUtils.getMockApplicationAttemptId(0, 0);
    FiCaSchedulerApp app_0 = new FiCaSchedulerApp(appAttemptId_0, user_0, a, mock(ActiveUsersManager.class), spyRMContext);
    app_0 = spy(app_0);
    Mockito.doNothing().when(app_0).updateAMContainerDiagnostics(any(AMState.class), any(String.class));
    rmContext.getRMApps().put(app_0.getApplicationId(), mock(RMApp.class));
    a.submitApplicationAttempt(app_0, user_0);
    final ApplicationAttemptId appAttemptId_1 = TestUtils.getMockApplicationAttemptId(1, 0);
    FiCaSchedulerApp app_1 = new FiCaSchedulerApp(appAttemptId_1, user_0, a, mock(ActiveUsersManager.class), spyRMContext);
    app_1 = spy(app_1);
    Mockito.doNothing().when(app_1).updateAMContainerDiagnostics(any(AMState.class), any(String.class));
    a.submitApplicationAttempt(app_1, user_0);
    // Setup some nodes
    String host_0 = "host_0";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, DEFAULT_RACK, 0, 8 * GB);
    String host_1 = "host_1";
    FiCaSchedulerNode node_1 = TestUtils.getMockNode(host_1, DEFAULT_RACK, 0, 8 * GB);
    Map<ApplicationAttemptId, FiCaSchedulerApp> apps = ImmutableMap.of(app_0.getApplicationAttemptId(), app_0, app_1.getApplicationAttemptId(), app_1);
    Map<NodeId, FiCaSchedulerNode> nodes = ImmutableMap.of(node_0.getNodeID(), node_0, node_1.getNodeID(), node_1);
    cs.getNodeTracker().addNode(node_0);
    cs.getNodeTracker().addNode(node_1);
    when(csContext.getNode(node_0.getNodeID())).thenReturn(node_0);
    when(csContext.getNode(node_1.getNodeID())).thenReturn(node_1);
    final int numNodes = 2;
    Resource clusterResource = Resources.createResource(numNodes * (8 * GB));
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Setup resource-requests
    Priority priorityAM = TestUtils.createMockPriority(1);
    Priority priorityMap = TestUtils.createMockPriority(5);
    Priority priorityReduce = TestUtils.createMockPriority(10);
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 2 * GB, 1, true, priorityAM, recordFactory)));
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 5 * GB, 2, true, priorityReduce, recordFactory)));
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 3 * GB, 2, true, priorityMap, recordFactory)));
    // Start testing...
    // Only AM
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(2 * GB, a.getUsedResources().getMemorySize());
    assertEquals(2 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, a.getMetrics().getReservedMB());
    assertEquals(2 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(14 * GB, a.getMetrics().getAvailableMB());
    assertEquals(2 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_1.getAllocatedResource().getMemorySize());
    // Only 1 map - simulating reduce
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(5 * GB, a.getUsedResources().getMemorySize());
    assertEquals(5 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, a.getMetrics().getReservedMB());
    assertEquals(5 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(11 * GB, a.getMetrics().getAvailableMB());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(0 * GB, node_1.getAllocatedResource().getMemorySize());
    // Only 1 map to other node - simulating reduce
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(8 * GB, a.getUsedResources().getMemorySize());
    assertEquals(8 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, a.getMetrics().getReservedMB());
    assertEquals(8 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(8 * GB, a.getMetrics().getAvailableMB());
    assertEquals(8 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(null, node_0.getReservedContainer());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(3 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(2, app_0.getOutstandingAsksCount(toSchedulerKey(priorityReduce)));
    // try to assign reducer (5G on node 0 and should reserve)
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(13 * GB, a.getUsedResources().getMemorySize());
    assertEquals(8 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(5 * GB, a.getMetrics().getReservedMB());
    assertEquals(8 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(3 * GB, a.getMetrics().getAvailableMB());
    assertEquals(3 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(5 * GB, node_0.getReservedContainer().getReservedResource().getMemorySize());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(3 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(2, app_0.getOutstandingAsksCount(toSchedulerKey(priorityReduce)));
    // could allocate but told need to unreserve first
    TestUtils.applyResourceCommitRequest(clusterResource, a.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), nodes, apps);
    assertEquals(13 * GB, a.getUsedResources().getMemorySize());
    assertEquals(13 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, a.getMetrics().getReservedMB());
    assertEquals(13 * GB, a.getMetrics().getAllocatedMB());
    assertEquals(3 * GB, a.getMetrics().getAvailableMB());
    assertEquals(3 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(null, node_0.getReservedContainer());
    assertEquals(5 * GB, node_0.getAllocatedResource().getMemorySize());
    assertEquals(8 * GB, node_1.getAllocatedResource().getMemorySize());
    assertEquals(1, app_0.getOutstandingAsksCount(toSchedulerKey(priorityReduce)));
}

Example 74 with FiCaSchedulerNode

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode in project hadoop by apache.

the class TestParentQueue method testOffSwitchScheduling.

@Test
public void testOffSwitchScheduling() throws Exception {
    // Setup queue configs
    setupSingleLevelQueues(csConf);
    Map<String, CSQueue> queues = new HashMap<String, CSQueue>();
    CSQueue root = CapacitySchedulerQueueManager.parseQueue(csContext, csConf, null, CapacitySchedulerConfiguration.ROOT, queues, queues, TestUtils.spyHook);
    // Setup some nodes
    final int memoryPerNode = 10;
    final int coresPerNode = 16;
    final int numNodes = 2;
    FiCaSchedulerNode node_0 = TestUtils.getMockNode("host_0", DEFAULT_RACK, 0, memoryPerNode * GB);
    FiCaSchedulerNode node_1 = TestUtils.getMockNode("host_1", DEFAULT_RACK, 0, memoryPerNode * GB);
    final Resource clusterResource = Resources.createResource(numNodes * (memoryPerNode * GB), numNodes * coresPerNode);
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Start testing
    LeafQueue a = (LeafQueue) queues.get(A);
    LeafQueue b = (LeafQueue) queues.get(B);
    a.getQueueResourceUsage().incPending(Resources.createResource(1 * GB));
    b.getQueueResourceUsage().incPending(Resources.createResource(1 * GB));
    queues.get(CapacitySchedulerConfiguration.ROOT).getQueueResourceUsage().incPending(Resources.createResource(1 * GB));
    // Simulate B returning a container on node_0
    stubQueueAllocation(a, clusterResource, node_0, 0 * GB, NodeType.OFF_SWITCH);
    stubQueueAllocation(b, clusterResource, node_0, 1 * GB, NodeType.OFF_SWITCH);
    root.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    verifyQueueMetrics(a, 0 * GB, clusterResource);
    verifyQueueMetrics(b, 1 * GB, clusterResource);
    // Now, A should get the scheduling opportunity since A=0G/6G, B=1G/14G
    // also, B gets a scheduling opportunity since A allocates RACK_LOCAL
    stubQueueAllocation(a, clusterResource, node_1, 2 * GB, NodeType.RACK_LOCAL);
    stubQueueAllocation(b, clusterResource, node_1, 1 * GB, NodeType.OFF_SWITCH);
    root.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    InOrder allocationOrder = inOrder(a);
    allocationOrder.verify(a).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    root.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    allocationOrder = inOrder(b);
    allocationOrder.verify(b).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    verifyQueueMetrics(a, 2 * GB, clusterResource);
    verifyQueueMetrics(b, 2 * GB, clusterResource);
    // Now, B should get the scheduling opportunity
    // since A has 2/6G while B has 2/14G,
    // However, since B returns off-switch, A won't get an opportunity
    stubQueueAllocation(a, clusterResource, node_0, 1 * GB, NodeType.NODE_LOCAL);
    stubQueueAllocation(b, clusterResource, node_0, 2 * GB, NodeType.OFF_SWITCH);
    root.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    allocationOrder = inOrder(b, a);
    allocationOrder.verify(b).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    allocationOrder.verify(a).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    verifyQueueMetrics(a, 2 * GB, clusterResource);
    verifyQueueMetrics(b, 4 * GB, clusterResource);
}

Example 75 with FiCaSchedulerNode

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode in project hadoop by apache.

the class TestParentQueue method testOffSwitchSchedulingMultiLevelQueues.

@Test
public void testOffSwitchSchedulingMultiLevelQueues() throws Exception {
    // Setup queue configs
    setupMultiLevelQueues(csConf);
    //B3
    Map<String, CSQueue> queues = new HashMap<String, CSQueue>();
    CSQueue root = CapacitySchedulerQueueManager.parseQueue(csContext, csConf, null, CapacitySchedulerConfiguration.ROOT, queues, queues, TestUtils.spyHook);
    // Setup some nodes
    final int memoryPerNode = 10;
    final int coresPerNode = 10;
    final int numNodes = 2;
    FiCaSchedulerNode node_0 = TestUtils.getMockNode("host_0", DEFAULT_RACK, 0, memoryPerNode * GB);
    FiCaSchedulerNode node_1 = TestUtils.getMockNode("host_1", DEFAULT_RACK, 0, memoryPerNode * GB);
    final Resource clusterResource = Resources.createResource(numNodes * (memoryPerNode * GB), numNodes * coresPerNode);
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Start testing
    LeafQueue b3 = (LeafQueue) queues.get(B3);
    LeafQueue b2 = (LeafQueue) queues.get(B2);
    b2.getQueueResourceUsage().incPending(Resources.createResource(1 * GB));
    b3.getQueueResourceUsage().incPending(Resources.createResource(1 * GB));
    queues.get(CapacitySchedulerConfiguration.ROOT).getQueueResourceUsage().incPending(Resources.createResource(1 * GB));
    CSQueue b = queues.get(B);
    b.getQueueResourceUsage().incPending(Resources.createResource(1 * GB));
    // Simulate B3 returning a container on node_0
    stubQueueAllocation(b2, clusterResource, node_0, 0 * GB, NodeType.OFF_SWITCH);
    stubQueueAllocation(b3, clusterResource, node_0, 1 * GB, NodeType.OFF_SWITCH);
    root.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    verifyQueueMetrics(b2, 0 * GB, clusterResource);
    verifyQueueMetrics(b3, 1 * GB, clusterResource);
    // Now, B2 should get the scheduling opportunity since B2=0G/2G, B3=1G/7G
    // also, B3 gets a scheduling opportunity since B2 allocates RACK_LOCAL
    stubQueueAllocation(b2, clusterResource, node_1, 1 * GB, NodeType.RACK_LOCAL);
    stubQueueAllocation(b3, clusterResource, node_1, 1 * GB, NodeType.OFF_SWITCH);
    root.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    root.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    InOrder allocationOrder = inOrder(b2, b3);
    allocationOrder.verify(b2).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    allocationOrder.verify(b3).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    verifyQueueMetrics(b2, 1 * GB, clusterResource);
    verifyQueueMetrics(b3, 2 * GB, clusterResource);
    // Now, B3 should get the scheduling opportunity 
    // since B2 has 1/2G while B3 has 2/7G, 
    // However, since B3 returns off-switch, B2 won't get an opportunity
    stubQueueAllocation(b2, clusterResource, node_0, 1 * GB, NodeType.NODE_LOCAL);
    stubQueueAllocation(b3, clusterResource, node_0, 1 * GB, NodeType.OFF_SWITCH);
    root.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    allocationOrder = inOrder(b3, b2);
    allocationOrder.verify(b3).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    allocationOrder.verify(b2).assignContainers(eq(clusterResource), any(PlacementSet.class), anyResourceLimits(), any(SchedulingMode.class));
    verifyQueueMetrics(b2, 1 * GB, clusterResource);
    verifyQueueMetrics(b3, 3 * GB, clusterResource);
}