Examples with NodeUpdateSchedulerEvent org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent used on opensource projects

Example 26 with NodeUpdateSchedulerEvent

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent in project hadoop by apache.

the class TestCapacitySchedulerSurgicalPreemption method testSurgicalPreemptionWithAvailableResource.

@Test(timeout = 60000)
public void testSurgicalPreemptionWithAvailableResource() throws Exception {
    /**
     * Test case: Submit two application (app1/app2) to different queues, queue
     * structure:
     *
     * <pre>
     *             Root
     *            /  |  \
     *           a   b   c
     *          10   20  70
     * </pre>
     *
     * 1) Two nodes (n1/n2) in the cluster, each of them has 20G.
     *
     * 2) app1 submit to queue-b, asks for 1G * 5
     *
     * 3) app2 submit to queue-c, ask for one 4G container (for AM)
     *
     * After preemption, we should expect:
     * Preempt 3 containers from app1 and AM of app2 successfully allocated.
     */
    MockRM rm1 = new MockRM(conf);
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 20 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 20 * GB);
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    RMNode rmNode2 = rm1.getRMContext().getRMNodes().get(nm2.getNodeId());
    // launch an app to queue, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "a");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    am1.allocate("*", 1 * GB, 38, new ArrayList<ContainerId>());
    // Do allocation for node1/node2
    for (int i = 0; i < 38; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
        cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    }
    // App1 should have 31 containers now
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    Assert.assertEquals(39, schedulerApp1.getLiveContainers().size());
    // 17 from n1 and 16 from n2
    waitNumberOfLiveContainersOnNodeFromApp(cs.getNode(rmNode1.getNodeID()), am1.getApplicationAttemptId(), 20);
    waitNumberOfLiveContainersOnNodeFromApp(cs.getNode(rmNode2.getNodeID()), am1.getApplicationAttemptId(), 19);
    // Submit app2 to queue-c and asks for a 4G container for AM
    RMApp app2 = rm1.submitApp(4 * GB, "app", "user", null, "c");
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(ApplicationAttemptId.newInstance(app2.getApplicationId(), 1));
    // Call editSchedule: containers are selected to be preemption candidate
    ProportionalCapacityPreemptionPolicy editPolicy = (ProportionalCapacityPreemptionPolicy) getSchedulingEditPolicy(rm1);
    editPolicy.editSchedule();
    Assert.assertEquals(3, editPolicy.getToPreemptContainers().size());
    // Call editSchedule again: selected containers are killed
    editPolicy.editSchedule();
    waitNumberOfLiveContainersFromApp(schedulerApp1, 36);
    // Call allocation, containers are reserved
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    waitNumberOfReservedContainersFromApp(schedulerApp2, 1);
    // Call editSchedule twice and allocation once, container should get allocated
    editPolicy.editSchedule();
    editPolicy.editSchedule();
    int tick = 0;
    while (schedulerApp2.getLiveContainers().size() != 1 && tick < 10) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
        cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
        tick++;
        Thread.sleep(100);
    }
    waitNumberOfReservedContainersFromApp(schedulerApp2, 0);
    rm1.close();
}

Example 27 with NodeUpdateSchedulerEvent

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent in project hadoop by apache.

the class TestContainerAllocation method testExcessReservationWillBeUnreserved.

@Test(timeout = 60000)
public void testExcessReservationWillBeUnreserved() throws Exception {
    /**
     * Test case: Submit two application (app1/app2) to a queue. And there's one
     * node with 8G resource in the cluster. App1 allocates a 6G container, Then
     * app2 asks for a 4G container. App2's request will be reserved on the
     * node.
     * 
     * Before next node heartbeat, app2 cancels the reservation, we should found
     * the reserved resource is cancelled as well.
     */
    // inject node label manager
    MockRM rm1 = new MockRM();
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 8 * GB);
    // launch an app to queue, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    // launch another app to queue, AM container should be launched in nm1
    RMApp app2 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm1);
    am1.allocate("*", 4 * GB, 1, new ArrayList<ContainerId>());
    am2.allocate("*", 4 * GB, 1, new ArrayList<ContainerId>());
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    LeafQueue leafQueue = (LeafQueue) cs.getQueue("default");
    // Do node heartbeats 2 times
    // First time will allocate container for app1, second time will reserve
    // container for app2
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    // App2 will get preference to be allocated on node1, and node1 will be all
    // used by App2.
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(am2.getApplicationAttemptId());
    // Check if a 4G contaienr allocated for app1, and nothing allocated for app2
    Assert.assertEquals(2, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp2.getLiveContainers().size());
    Assert.assertTrue(schedulerApp2.getReservedContainers().size() > 0);
    // NM1 has available resource = 2G (8G - 2 * 1G - 4G)
    Assert.assertEquals(2 * GB, cs.getNode(nm1.getNodeId()).getUnallocatedResource().getMemorySize());
    Assert.assertNotNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Usage of queue = 4G + 2 * 1G + 4G (reserved)
    Assert.assertEquals(10 * GB, cs.getRootQueue().getQueueResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(4 * GB, cs.getRootQueue().getQueueResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(4 * GB, leafQueue.getQueueResourceUsage().getReserved().getMemorySize());
    // Cancel asks of app2 and re-kick RM
    am2.allocate("*", 4 * GB, 0, new ArrayList<ContainerId>());
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    // App2's reservation will be cancelled
    Assert.assertTrue(schedulerApp2.getReservedContainers().size() == 0);
    Assert.assertEquals(2 * GB, cs.getNode(nm1.getNodeId()).getUnallocatedResource().getMemorySize());
    Assert.assertNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    Assert.assertEquals(6 * GB, cs.getRootQueue().getQueueResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(0, cs.getRootQueue().getQueueResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(0, leafQueue.getQueueResourceUsage().getReserved().getMemorySize());
    rm1.close();
}

Example 28 with NodeUpdateSchedulerEvent

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent in project hadoop by apache.

the class TestContainerAllocation method testQueuePriorityOrdering.

@Test(timeout = 60000)
public void testQueuePriorityOrdering() throws Exception {
    CapacitySchedulerConfiguration newConf = (CapacitySchedulerConfiguration) TestUtils.getConfigurationWithMultipleQueues(conf);
    // Set ordering policy
    newConf.setQueueOrderingPolicy(CapacitySchedulerConfiguration.ROOT, CapacitySchedulerConfiguration.QUEUE_PRIORITY_UTILIZATION_ORDERING_POLICY);
    // Set maximum capacity of A to 20
    newConf.setMaximumCapacity(CapacitySchedulerConfiguration.ROOT + ".a", 20);
    newConf.setQueuePriority(CapacitySchedulerConfiguration.ROOT + ".c", 1);
    newConf.setQueuePriority(CapacitySchedulerConfiguration.ROOT + ".b", 2);
    newConf.setQueuePriority(CapacitySchedulerConfiguration.ROOT + ".a", 3);
    MockRM rm1 = new MockRM(newConf);
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 100 * GB);
    // launch an app to queue A, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(2 * GB, "app", "user", null, "a");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    // launch an app to queue B, AM container should be launched in nm1
    RMApp app2 = rm1.submitApp(2 * GB, "app", "user", null, "b");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm1);
    // launch an app to queue C, AM container should be launched in nm1
    RMApp app3 = rm1.submitApp(2 * GB, "app", "user", null, "c");
    MockAM am3 = MockRM.launchAndRegisterAM(app3, rm1, nm1);
    // Each application asks 10 * 5GB containers
    am1.allocate("*", 5 * GB, 10, null);
    am2.allocate("*", 5 * GB, 10, null);
    am3.allocate("*", 5 * GB, 10, null);
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(am2.getApplicationAttemptId());
    FiCaSchedulerApp schedulerApp3 = cs.getApplicationAttempt(am3.getApplicationAttemptId());
    // container will be allocated to am1
    // App1 will get 2 container allocated (plus AM container)
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    Assert.assertEquals(2, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp2.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp3.getLiveContainers().size());
    // container will be allocated to am1 again,
    // App1 will get 3 container allocated (plus AM container)
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    Assert.assertEquals(3, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp2.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp3.getLiveContainers().size());
    // (Now usages of queues: a=12G (satisfied), b=2G, c=2G)
    // container will be allocated to am2 (since app1 reaches its guaranteed
    // capacity)
    // App2 will get 2 container allocated (plus AM container)
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    Assert.assertEquals(3, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(2, schedulerApp2.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp3.getLiveContainers().size());
    // App2 will get 2 container allocated (plus AM container)
    for (int i = 0; i < 3; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    }
    Assert.assertEquals(3, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(5, schedulerApp2.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp3.getLiveContainers().size());
    // Do this 10 times
    for (int i = 0; i < 10; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    }
    Assert.assertEquals(3, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(5, schedulerApp2.getLiveContainers().size());
    Assert.assertEquals(11, schedulerApp3.getLiveContainers().size());
    // Do this 20 times, we can only allocate 3 containers, 1 to A and 3 to B
    for (int i = 0; i < 20; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    }
    Assert.assertEquals(4, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(6, schedulerApp2.getLiveContainers().size());
    Assert.assertEquals(11, schedulerApp3.getLiveContainers().size());
    // (Now usages of queues: a=17G (satisfied), b=27G (satisfied), c=52G))
    rm1.close();
}

Example 29 with NodeUpdateSchedulerEvent

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent in project hadoop by apache.

the class TestContainerAllocation method testReservedContainerMetricsOnDecommisionedNode.

@Test(timeout = 60000)
public void testReservedContainerMetricsOnDecommisionedNode() throws Exception {
    /**
     * Test case: Submit two application (app1/app2) to a queue. And there's one
     * node with 8G resource in the cluster. App1 allocates a 6G container, Then
     * app2 asks for a 4G container. App2's request will be reserved on the
     * node.
     *
     * Before next node heartbeat, app1 container is completed/killed. So app1
     * container which was reserved will be allocated.
     */
    // inject node label manager
    MockRM rm1 = new MockRM();
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 8 * GB);
    // launch an app to queue, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    // launch another app to queue, AM container should be launched in nm1
    RMApp app2 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm1);
    am1.allocate("*", 4 * GB, 1, new ArrayList<ContainerId>());
    am2.allocate("*", 4 * GB, 1, new ArrayList<ContainerId>());
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    LeafQueue leafQueue = (LeafQueue) cs.getQueue("default");
    // Do node heartbeats 2 times
    // First time will allocate container for app1, second time will reserve
    // container for app2
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    // App2 will get preference to be allocated on node1, and node1 will be all
    // used by App2.
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(am2.getApplicationAttemptId());
    // Check if a 4G container allocated for app1, and nothing allocated for app2
    Assert.assertEquals(2, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp2.getLiveContainers().size());
    Assert.assertTrue(schedulerApp2.getReservedContainers().size() > 0);
    // NM1 has available resource = 2G (8G - 2 * 1G - 4G)
    Assert.assertEquals(2 * GB, cs.getNode(nm1.getNodeId()).getUnallocatedResource().getMemorySize());
    Assert.assertNotNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Usage of queue = 4G + 2 * 1G + 4G (reserved)
    Assert.assertEquals(10 * GB, cs.getRootQueue().getQueueResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(4 * GB, cs.getRootQueue().getQueueResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(4 * GB, leafQueue.getQueueResourceUsage().getReserved().getMemorySize());
    // Remove the node
    cs.handle(new NodeRemovedSchedulerEvent(rmNode1));
    // Check all container cancelled for app1 and app2
    Assert.assertEquals(0, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(0, schedulerApp2.getLiveContainers().size());
    Assert.assertFalse(schedulerApp2.getReservedContainers().size() > 0);
    // Usage and Reserved capacity of queue is 0
    Assert.assertEquals(0 * GB, cs.getRootQueue().getQueueResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(0 * GB, cs.getRootQueue().getQueueResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(0 * GB, leafQueue.getQueueResourceUsage().getReserved().getMemorySize());
    rm1.close();
}

Example 30 with NodeUpdateSchedulerEvent

use of org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent in project hadoop by apache.

the class TestContainerAllocation method testAssignMultipleOffswitchContainers.

@Test(timeout = 60000)
public void testAssignMultipleOffswitchContainers() throws Exception {
    MockRM rm1 = new MockRM();
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 80 * GB);
    // launch an app to queue, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    am1.allocate("*", 1 * GB, 5, new ArrayList<ContainerId>());
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    // Do node heartbeats once
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    // App1 will get one container allocated (plus AM container
    Assert.assertEquals(2, schedulerApp1.getLiveContainers().size());
    // Set assign multiple off-switch containers to 3
    CapacitySchedulerConfiguration newCSConf = new CapacitySchedulerConfiguration();
    newCSConf.setInt(CapacitySchedulerConfiguration.OFFSWITCH_PER_HEARTBEAT_LIMIT, 3);
    cs.reinitialize(newCSConf, rm1.getRMContext());
    // Do node heartbeats once
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    // App1 will get 3 new container allocated (plus 2 previously allocated
    // container)
    Assert.assertEquals(5, schedulerApp1.getLiveContainers().size());
    rm1.close();
}