Search in sources :

Example 46 with FiCaSchedulerApp

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

the class TestIncreaseAllocationExpirer method testContainerIncreaseAllocationExpiration.

@Test
public void testContainerIncreaseAllocationExpiration() throws Exception {
    /**
     * 1. Allocate 1 container: containerId2 (1G)
     * 2. Increase resource of containerId2: 1G -> 3G
     * 3. AM acquires the token
     * 4. AM does not use the token
     * 5. Verify containerId2's resource usage falls back to
     *    1G after the increase token expires
     */
    // Set the allocation expiration to 5 seconds
    conf.setLong(YarnConfiguration.RM_CONTAINER_ALLOC_EXPIRY_INTERVAL_MS, 5000);
    MockRM rm1 = new MockRM(conf);
    rm1.start();
    MockNM nm1 = rm1.registerNode("127.0.0.1:1234", 20 * GB);
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    nm1.nodeHeartbeat(app1.getCurrentAppAttempt().getAppAttemptId(), 1, ContainerState.RUNNING);
    am1.allocate("127.0.0.1", 1 * GB, 1, new ArrayList<ContainerId>());
    ContainerId containerId2 = ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
    rm1.waitForState(nm1, containerId2, RMContainerState.ALLOCATED);
    List<Container> containers = am1.allocate(null, null).getAllocatedContainers();
    Assert.assertEquals(containerId2, containers.get(0).getId());
    Assert.assertNotNull(containers.get(0).getContainerToken());
    checkUsedResource(rm1, "default", 2 * GB, null);
    FiCaSchedulerApp app = TestUtils.getFiCaSchedulerApp(rm1, app1.getApplicationId());
    Assert.assertEquals(2 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    verifyAvailableResourceOfSchedulerNode(rm1, nm1.getNodeId(), 18 * GB);
    nm1.nodeHeartbeat(app1.getCurrentAppAttempt().getAppAttemptId(), 2, ContainerState.RUNNING);
    rm1.waitForState(nm1, containerId2, RMContainerState.RUNNING);
    // am1 asks to increase containerId2 from 1GB to 3GB
    am1.sendContainerResizingRequest(Collections.singletonList(UpdateContainerRequest.newInstance(0, containerId2, ContainerUpdateType.INCREASE_RESOURCE, Resources.createResource(3 * GB), null)));
    // Kick off scheduling and wait for 1 second;
    nm1.nodeHeartbeat(true);
    Thread.sleep(1000);
    // Start container increase allocation expirer
    am1.allocate(null, null);
    // Verify resource usage
    checkUsedResource(rm1, "default", 4 * GB, null);
    Assert.assertEquals(4 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    verifyAvailableResourceOfSchedulerNode(rm1, nm1.getNodeId(), 16 * GB);
    // Wait long enough for the increase token to expire, and for the roll
    // back action to complete
    Thread.sleep(10000);
    // Verify container size is 1G
    am1.allocate(null, null);
    Assert.assertEquals(1 * GB, rm1.getResourceScheduler().getRMContainer(containerId2).getAllocatedResource().getMemorySize());
    // Verify total resource usage is 2G
    checkUsedResource(rm1, "default", 2 * GB, null);
    Assert.assertEquals(2 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    // Verify available resource is rolled back to 18GB
    verifyAvailableResourceOfSchedulerNode(rm1, nm1.getNodeId(), 18 * GB);
    rm1.stop();
}
Also used : RMApp(org.apache.hadoop.yarn.server.resourcemanager.rmapp.RMApp) RMContainer(org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainer) Container(org.apache.hadoop.yarn.api.records.Container) ContainerId(org.apache.hadoop.yarn.api.records.ContainerId) MockNM(org.apache.hadoop.yarn.server.resourcemanager.MockNM) FiCaSchedulerApp(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp) MockAM(org.apache.hadoop.yarn.server.resourcemanager.MockAM) MockRM(org.apache.hadoop.yarn.server.resourcemanager.MockRM) Test(org.junit.Test)

Example 47 with FiCaSchedulerApp

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

the class TestLeafQueue method testSingleQueueOneUserMetrics.

@Test
public void testSingleQueueOneUserMetrics() throws Exception {
    // Manipulate queue 'a'
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(B));
    // 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);
    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);
    // same user
    a.submitApplicationAttempt(app_1, user_0);
    // Setup some nodes
    String host_0 = "127.0.0.1";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, 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);
    final int numNodes = 1;
    Resource clusterResource = Resources.createResource(numNodes * (8 * GB), numNodes * 16);
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Setup resource-requests
    Priority priority = TestUtils.createMockPriority(1);
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 3, true, priority, recordFactory)));
    // Start testing...
    // Only 1 container
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals((int) (node_0.getTotalResource().getMemorySize() * a.getCapacity()) - (1 * GB), a.getMetrics().getAvailableMB());
}
Also used : FiCaSchedulerNode(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode) Priority(org.apache.hadoop.yarn.api.records.Priority) Resource(org.apache.hadoop.yarn.api.records.Resource) ApplicationAttemptId(org.apache.hadoop.yarn.api.records.ApplicationAttemptId) FiCaSchedulerApp(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp) ResourceLimits(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ResourceLimits) NodeId(org.apache.hadoop.yarn.api.records.NodeId) ActiveUsersManager(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ActiveUsersManager) Test(org.junit.Test)

Example 48 with FiCaSchedulerApp

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

the class TestLeafQueue method testDRFUsageRatioRounding.

@Test
public void testDRFUsageRatioRounding() throws Exception {
    CSAssignment assign;
    setUpWithDominantResourceCalculator();
    // Mock the queue
    LeafQueue b = stubLeafQueue((LeafQueue) queues.get(E));
    // 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, 80 * GB, 100);
    // Make cluster relatively large so usageRatios are small
    int numNodes = 1000;
    Resource clusterResource = Resources.createResource(numNodes * (80 * GB), numNodes * 100);
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Set user-limit. Need a small queue within a large cluster.
    b.setUserLimit(50);
    b.setUserLimitFactor(1000000);
    b.setMaxCapacity(1.0f);
    b.setAbsoluteCapacity(0.00001f);
    // First allocation is larger than second but is still vcore dominant
    // so usage ratio will be based on vcores. If consumedRatio doesn't round
    // in our favor then new limit calculation will actually be less than
    // what is currently consumed and we will fail to allocate
    Priority priority = TestUtils.createMockPriority(1);
    app0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 20 * GB, 29, 1, true, priority, recordFactory, RMNodeLabelsManager.NO_LABEL)));
    assign = b.assignContainers(clusterResource, node0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    app0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 10 * GB, 29, 2, true, priority, recordFactory, RMNodeLabelsManager.NO_LABEL)));
    assign = b.assignContainers(clusterResource, node0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    assertTrue("Still within limits, should assign", assign.getResource().getMemorySize() > 0);
}
Also used : FiCaSchedulerNode(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode) Priority(org.apache.hadoop.yarn.api.records.Priority) FiCaSchedulerApp(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp) ResourceLimits(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ResourceLimits) Resource(org.apache.hadoop.yarn.api.records.Resource) ApplicationAttemptId(org.apache.hadoop.yarn.api.records.ApplicationAttemptId) Test(org.junit.Test)

Example 49 with FiCaSchedulerApp

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

the class TestLeafQueue method testLocalityScheduling.

@Test
public void testLocalityScheduling() throws Exception {
    // Manipulate queue 'b'
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(B));
    // User
    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);
    a.submitApplicationAttempt(app_0, user_0);
    // Setup some nodes and racks
    String host_0 = "127.0.0.1";
    String rack_0 = "rack_0";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, rack_0, 0, 8 * GB);
    String host_1 = "127.0.0.2";
    String rack_1 = "rack_1";
    FiCaSchedulerNode node_1 = TestUtils.getMockNode(host_1, rack_1, 0, 8 * GB);
    String host_2 = "127.0.0.3";
    String rack_2 = "rack_2";
    FiCaSchedulerNode node_2 = TestUtils.getMockNode(host_2, rack_2, 0, 8 * GB);
    // on rack_1
    String host_3 = "127.0.0.4";
    FiCaSchedulerNode node_3 = TestUtils.getMockNode(host_3, rack_1, 0, 8 * GB);
    Map<ApplicationAttemptId, FiCaSchedulerApp> apps = ImmutableMap.of(app_0.getApplicationAttemptId(), app_0);
    Map<NodeId, FiCaSchedulerNode> nodes = ImmutableMap.of(node_0.getNodeID(), node_0, node_1.getNodeID(), node_1, node_2.getNodeID(), node_2, node_3.getNodeID(), node_3);
    final int numNodes = 3;
    Resource clusterResource = Resources.createResource(numNodes * (8 * GB), numNodes * 16);
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Setup resource-requests and submit
    Priority priority = TestUtils.createMockPriority(1);
    List<ResourceRequest> app_0_requests_0 = new ArrayList<ResourceRequest>();
    app_0_requests_0.add(TestUtils.createResourceRequest(host_0, 1 * GB, 1, true, priority, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(rack_0, 1 * GB, 1, true, priority, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(host_1, 1 * GB, 1, true, priority, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(rack_1, 1 * GB, 1, true, priority, recordFactory));
    app_0_requests_0.add(// one extra
    TestUtils.createResourceRequest(// one extra
    ResourceRequest.ANY, // one extra
    1 * GB, // one extra
    3, true, priority, recordFactory));
    app_0.updateResourceRequests(app_0_requests_0);
    // Start testing...
    CSAssignment assignment = null;
    SchedulerRequestKey schedulerKey = toSchedulerKey(priority);
    // Start with off switch, shouldn't allocate due to delay scheduling
    assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyNoContainerAllocated(assignment);
    assertEquals(1, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(3, app_0.getOutstandingAsksCount(schedulerKey));
    // None->NODE_LOCAL
    assertEquals(NodeType.NODE_LOCAL, assignment.getType());
    // Another off switch, shouldn't allocate due to delay scheduling
    assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyNoContainerAllocated(assignment);
    assertEquals(2, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(3, app_0.getOutstandingAsksCount(schedulerKey));
    // None->NODE_LOCAL
    assertEquals(NodeType.NODE_LOCAL, assignment.getType());
    // Another off switch, shouldn't allocate due to delay scheduling
    assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyNoContainerAllocated(assignment);
    assertEquals(3, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(3, app_0.getOutstandingAsksCount(schedulerKey));
    // None->NODE_LOCAL
    assertEquals(NodeType.NODE_LOCAL, assignment.getType());
    // Another off switch, now we should allocate 
    // since missedOpportunities=3 and reqdContainers=3
    assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.OFF_SWITCH);
    // should NOT reset
    assertEquals(4, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(2, app_0.getOutstandingAsksCount(schedulerKey));
    // NODE_LOCAL - node_0
    assignment = a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.NODE_LOCAL);
    // should reset
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(1, app_0.getOutstandingAsksCount(schedulerKey));
    // NODE_LOCAL - node_1
    assignment = a.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.NODE_LOCAL);
    // should reset
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(0, app_0.getOutstandingAsksCount(schedulerKey));
    assertEquals(NodeType.NODE_LOCAL, assignment.getType());
    // Add 1 more request to check for RACK_LOCAL
    app_0_requests_0.clear();
    app_0_requests_0.add(TestUtils.createResourceRequest(host_1, 1 * GB, 1, true, priority, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(rack_1, 1 * GB, 3, true, priority, recordFactory));
    app_0_requests_0.add(// one extra
    TestUtils.createResourceRequest(// one extra
    ResourceRequest.ANY, // one extra
    1 * GB, // one extra
    4, true, priority, recordFactory));
    app_0.updateResourceRequests(app_0_requests_0);
    assertEquals(4, app_0.getOutstandingAsksCount(schedulerKey));
    // Rack-delay
    doReturn(true).when(a).getRackLocalityFullReset();
    doReturn(1).when(a).getNodeLocalityDelay();
    // Shouldn't assign RACK_LOCAL yet
    assignment = a.assignContainers(clusterResource, node_3, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    assertEquals(1, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(4, app_0.getOutstandingAsksCount(schedulerKey));
    // Should assign RACK_LOCAL now
    assignment = a.assignContainers(clusterResource, node_3, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.RACK_LOCAL);
    // should reset
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(3, app_0.getOutstandingAsksCount(schedulerKey));
    // Shouldn't assign RACK_LOCAL because schedulingOpportunities should have gotten reset.
    assignment = a.assignContainers(clusterResource, node_3, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    assertEquals(1, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(3, app_0.getOutstandingAsksCount(schedulerKey));
    // Next time we schedule RACK_LOCAL, don't reset
    doReturn(false).when(a).getRackLocalityFullReset();
    // Should assign RACK_LOCAL now
    assignment = a.assignContainers(clusterResource, node_3, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.RACK_LOCAL);
    // should NOT reset
    assertEquals(2, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(2, app_0.getOutstandingAsksCount(schedulerKey));
    // Another RACK_LOCAL since schedulingOpportunities not reset
    assignment = a.assignContainers(clusterResource, node_3, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.RACK_LOCAL);
    // should NOT reset
    assertEquals(3, app_0.getSchedulingOpportunities(schedulerKey));
    assertEquals(1, app_0.getOutstandingAsksCount(schedulerKey));
    // Add a request larger than cluster size to verify
    // OFF_SWITCH delay is capped by cluster size
    app_0.resetSchedulingOpportunities(schedulerKey);
    app_0_requests_0.clear();
    app_0_requests_0.add(TestUtils.createResourceRequest(host_0, 1 * GB, 100, true, priority, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(rack_0, 1 * GB, 100, true, priority, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 100, true, priority, recordFactory));
    app_0.updateResourceRequests(app_0_requests_0);
    // Start with off switch. 3 nodes in cluster so shouldn't allocate first 3
    for (int i = 0; i < numNodes; i++) {
        assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
        applyCSAssignment(clusterResource, assignment, a, nodes, apps);
        verifyNoContainerAllocated(assignment);
        assertEquals(i + 1, app_0.getSchedulingOpportunities(schedulerKey));
    }
    // delay should be capped at numNodes so next one should allocate
    assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.OFF_SWITCH);
    assertEquals(numNodes + 1, app_0.getSchedulingOpportunities(schedulerKey));
}
Also used : FiCaSchedulerNode(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode) Priority(org.apache.hadoop.yarn.api.records.Priority) Resource(org.apache.hadoop.yarn.api.records.Resource) ArrayList(java.util.ArrayList) ApplicationAttemptId(org.apache.hadoop.yarn.api.records.ApplicationAttemptId) SchedulerRequestKey(org.apache.hadoop.yarn.server.scheduler.SchedulerRequestKey) FiCaSchedulerApp(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp) ResourceLimits(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ResourceLimits) NodeId(org.apache.hadoop.yarn.api.records.NodeId) ActiveUsersManager(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ActiveUsersManager) ResourceRequest(org.apache.hadoop.yarn.api.records.ResourceRequest) Test(org.junit.Test)

Example 50 with FiCaSchedulerApp

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

the class TestLeafQueue method testApplicationPriorityScheduling.

@Test
public void testApplicationPriorityScheduling() throws Exception {
    // Manipulate queue 'a'
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(A));
    // User
    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);
    a.submitApplicationAttempt(app_0, user_0);
    // Setup some nodes and racks
    String host_0 = "127.0.0.1";
    String rack_0 = "rack_0";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, rack_0, 0, 8 * GB);
    String host_1 = "127.0.0.2";
    String rack_1 = "rack_1";
    FiCaSchedulerNode node_1 = TestUtils.getMockNode(host_1, rack_1, 0, 8 * GB);
    String host_2 = "127.0.0.3";
    String rack_2 = "rack_2";
    FiCaSchedulerNode node_2 = TestUtils.getMockNode(host_2, rack_2, 0, 8 * GB);
    Map<ApplicationAttemptId, FiCaSchedulerApp> apps = ImmutableMap.of(app_0.getApplicationAttemptId(), app_0);
    Map<NodeId, FiCaSchedulerNode> nodes = ImmutableMap.of(node_0.getNodeID(), node_0, node_1.getNodeID(), node_1, node_2.getNodeID(), node_2);
    final int numNodes = 3;
    Resource clusterResource = Resources.createResource(numNodes * (8 * GB), 1);
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    // Setup resource-requests and submit
    List<ResourceRequest> app_0_requests_0 = new ArrayList<ResourceRequest>();
    // P1
    Priority priority_1 = TestUtils.createMockPriority(1);
    SchedulerRequestKey schedulerKey1 = toSchedulerKey(priority_1);
    app_0_requests_0.add(TestUtils.createResourceRequest(host_0, 1 * GB, 1, true, priority_1, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(rack_0, 1 * GB, 1, true, priority_1, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(host_1, 1 * GB, 1, true, priority_1, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(rack_1, 1 * GB, 1, true, priority_1, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority_1, recordFactory));
    // P2
    Priority priority_2 = TestUtils.createMockPriority(2);
    SchedulerRequestKey schedulerKey2 = toSchedulerKey(priority_2);
    app_0_requests_0.add(TestUtils.createResourceRequest(host_2, 2 * GB, 1, true, priority_2, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(rack_2, 2 * GB, 1, true, priority_2, recordFactory));
    app_0_requests_0.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 2 * GB, 1, true, priority_2, recordFactory));
    app_0.updateResourceRequests(app_0_requests_0);
    // Start testing...
    // Start with off switch, shouldn't allocate P1 due to delay scheduling
    // thus, no P2 either!
    CSAssignment assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyNoContainerAllocated(assignment);
    assertEquals(1, app_0.getSchedulingOpportunities(schedulerKey1));
    assertEquals(2, app_0.getOutstandingAsksCount(schedulerKey1));
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey2));
    assertEquals(1, app_0.getOutstandingAsksCount(schedulerKey2));
    // Another off-switch, shouldn't allocate P1 due to delay scheduling
    // thus, no P2 either!
    assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyNoContainerAllocated(assignment);
    assertEquals(2, app_0.getSchedulingOpportunities(schedulerKey1));
    assertEquals(2, app_0.getOutstandingAsksCount(schedulerKey1));
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey2));
    assertEquals(1, app_0.getOutstandingAsksCount(schedulerKey2));
    // Another off-switch, shouldn't allocate OFF_SWITCH P1
    assignment = a.assignContainers(clusterResource, node_2, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.OFF_SWITCH);
    assertEquals(3, app_0.getSchedulingOpportunities(schedulerKey1));
    assertEquals(1, app_0.getOutstandingAsksCount(schedulerKey1));
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey2));
    assertEquals(1, app_0.getOutstandingAsksCount(schedulerKey2));
    // Now, DATA_LOCAL for P1
    assignment = a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.NODE_LOCAL);
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey1));
    assertEquals(0, app_0.getOutstandingAsksCount(schedulerKey1));
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey2));
    assertEquals(1, app_0.getOutstandingAsksCount(schedulerKey2));
    // Now, OFF_SWITCH for P2
    assignment = a.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
    applyCSAssignment(clusterResource, assignment, a, nodes, apps);
    verifyContainerAllocated(assignment, NodeType.OFF_SWITCH);
    assertEquals(0, app_0.getSchedulingOpportunities(schedulerKey1));
    assertEquals(0, app_0.getOutstandingAsksCount(schedulerKey1));
    assertEquals(1, app_0.getSchedulingOpportunities(schedulerKey2));
    assertEquals(0, app_0.getOutstandingAsksCount(schedulerKey2));
}
Also used : FiCaSchedulerNode(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode) Priority(org.apache.hadoop.yarn.api.records.Priority) Resource(org.apache.hadoop.yarn.api.records.Resource) ArrayList(java.util.ArrayList) ApplicationAttemptId(org.apache.hadoop.yarn.api.records.ApplicationAttemptId) SchedulerRequestKey(org.apache.hadoop.yarn.server.scheduler.SchedulerRequestKey) FiCaSchedulerApp(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp) ResourceLimits(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ResourceLimits) NodeId(org.apache.hadoop.yarn.api.records.NodeId) ActiveUsersManager(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ActiveUsersManager) ResourceRequest(org.apache.hadoop.yarn.api.records.ResourceRequest) Test(org.junit.Test)

Aggregations

FiCaSchedulerApp (org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp)141 Test (org.junit.Test)97 RMApp (org.apache.hadoop.yarn.server.resourcemanager.rmapp.RMApp)60 Resource (org.apache.hadoop.yarn.api.records.Resource)53 MockRM (org.apache.hadoop.yarn.server.resourcemanager.MockRM)51 ApplicationAttemptId (org.apache.hadoop.yarn.api.records.ApplicationAttemptId)49 MockNM (org.apache.hadoop.yarn.server.resourcemanager.MockNM)48 MockAM (org.apache.hadoop.yarn.server.resourcemanager.MockAM)47 FiCaSchedulerNode (org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerNode)47 ContainerId (org.apache.hadoop.yarn.api.records.ContainerId)40 Priority (org.apache.hadoop.yarn.api.records.Priority)40 RMContainer (org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainer)35 ResourceLimits (org.apache.hadoop.yarn.server.resourcemanager.scheduler.ResourceLimits)34 NodeId (org.apache.hadoop.yarn.api.records.NodeId)31 NodeUpdateSchedulerEvent (org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent)31 RMNode (org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode)30 ActiveUsersManager (org.apache.hadoop.yarn.server.resourcemanager.scheduler.ActiveUsersManager)24 ArrayList (java.util.ArrayList)19 YarnConfiguration (org.apache.hadoop.yarn.conf.YarnConfiguration)19 Container (org.apache.hadoop.yarn.api.records.Container)13