Examples with RMNode org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode used on opensource projects

Example 56 with RMNode

use of org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode 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();
}

Example 57 with RMNode

use of org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode in project hadoop by apache.

the class TestCapacitySchedulerLazyPreemption method testPreemptionConsidersNodeLocalityDelay.

@Test(timeout = 60000)
public void testPreemptionConsidersNodeLocalityDelay() throws Exception {
    /**
     * Test case: same as testSimplePreemption steps 1-3.
     *
     * Step 4: app2 asks for 1G container with locality specified, so it needs
     * to wait for missed-opportunity before get scheduled.
     * Check if system waits missed-opportunity before finish killable container
     */
    MockRM rm1 = new MockRM(conf);
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 4 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 4 * 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, 6, new ArrayList<ContainerId>());
    // Do allocation 3 times for node1/node2
    for (int i = 0; i < 3; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
        cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    }
    // App1 should have 7 containers now, and no available resource for cluster
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    Assert.assertEquals(7, schedulerApp1.getLiveContainers().size());
    // Submit app2 to queue-c and asks for a 1G container for AM
    RMApp app2 = rm1.submitApp(1 * GB, "app", "user", null, "c");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm2);
    // NM1/NM2 has available resource = 0G
    Assert.assertEquals(0 * GB, cs.getNode(nm1.getNodeId()).getUnallocatedResource().getMemorySize());
    Assert.assertEquals(0 * GB, cs.getNode(nm2.getNodeId()).getUnallocatedResource().getMemorySize());
    // AM asks for a 1 * GB container with unknown host and unknown rack
    am2.allocate(Arrays.asList(ResourceRequest.newInstance(Priority.newInstance(1), ResourceRequest.ANY, Resources.createResource(1 * GB), 1), ResourceRequest.newInstance(Priority.newInstance(1), "unknownhost", Resources.createResource(1 * GB), 1), ResourceRequest.newInstance(Priority.newInstance(1), "/default-rack", Resources.createResource(1 * GB), 1)), null);
    // Get edit policy and do one update
    SchedulingEditPolicy editPolicy = getSchedulingEditPolicy(rm1);
    // Call edit schedule twice, and check if one container from app1 marked
    // to be "killable"
    editPolicy.editSchedule();
    editPolicy.editSchedule();
    PreemptionManager pm = cs.getPreemptionManager();
    Map<ContainerId, RMContainer> killableContainers = waitKillableContainersSize(pm, "a", RMNodeLabelsManager.NO_LABEL, 1);
    Assert.assertEquals(killableContainers.entrySet().iterator().next().getKey().getApplicationAttemptId(), am1.getApplicationAttemptId());
    // Call CS.handle once to see if container preempted
    cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(am2.getApplicationAttemptId());
    // App1 has 7 containers, and app2 has 1 containers (no container preempted)
    Assert.assertEquals(7, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp2.getLiveContainers().size());
    // Do allocation again, one container will be preempted
    cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    // App1 has 6 containers, and app2 has 2 containers (new container allocated)
    Assert.assertEquals(6, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(2, schedulerApp2.getLiveContainers().size());
    rm1.close();
}

Example 58 with RMNode

use of org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode in project hadoop by apache.

the class TestCapacitySchedulerLazyPreemption method testPreemptionPolicyShouldRespectAlreadyMarkedKillableContainers.

@Test(timeout = 60000)
public void testPreemptionPolicyShouldRespectAlreadyMarkedKillableContainers() throws Exception {
    /**
     * Test case:
     * <pre>
     *             Root
     *            /  |  \
     *           a   b   c
     *          10   20  70
     * </pre>
     * Submit applications to two queues, one uses more than the other, so
     * preemption will happen.
     *
     * Check:
     * 1) Killable containers resources will be excluded from PCPP (no duplicated
     *    container added to killable list)
     * 2) When more resources need to be preempted, new containers will be selected
     *    and killable containers will be considered
     */
    MockRM rm1 = new MockRM(conf);
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.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, 6, new ArrayList<ContainerId>());
    // Do allocation 6 times for node1
    for (int i = 0; i < 6; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    }
    // App1 should have 7 containers now, and no available resource for cluster
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    Assert.assertEquals(7, schedulerApp1.getLiveContainers().size());
    // Submit app2 to queue-c and asks for a 1G container for AM
    RMApp app2 = rm1.submitApp(1 * GB, "app", "user", null, "c");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm1);
    // NM1 has available resource = 0G
    Assert.assertEquals(0 * GB, cs.getNode(nm1.getNodeId()).getUnallocatedResource().getMemorySize());
    am2.allocate("*", 1 * GB, 1, new ArrayList<ContainerId>());
    // Get edit policy and do one update
    ProportionalCapacityPreemptionPolicy editPolicy = (ProportionalCapacityPreemptionPolicy) getSchedulingEditPolicy(rm1);
    // Call edit schedule twice, and check if one container from app1 marked
    // to be "killable"
    editPolicy.editSchedule();
    editPolicy.editSchedule();
    PreemptionManager pm = cs.getPreemptionManager();
    waitKillableContainersSize(pm, "a", RMNodeLabelsManager.NO_LABEL, 1);
    // Check killable containers and to-be-preempted containers in edit policy
    Assert.assertEquals(0, editPolicy.getToPreemptContainers().size());
    // Run edit schedule again, confirm status doesn't changed
    editPolicy.editSchedule();
    Assert.assertEquals(0, editPolicy.getToPreemptContainers().size());
    // Save current to kill containers
    Set<ContainerId> previousKillableContainers = new HashSet<>(pm.getKillableContainersMap("a", RMNodeLabelsManager.NO_LABEL).keySet());
    // Update request resource of c from 1 to 2, so we need to preempt
    // one more container
    am2.allocate("*", 1 * GB, 2, new ArrayList<ContainerId>());
    // Call editPolicy.editSchedule() once, we should have 1 container in to-preempt map
    // and 1 container in killable map
    editPolicy.editSchedule();
    Assert.assertEquals(1, editPolicy.getToPreemptContainers().size());
    // Call editPolicy.editSchedule() once more, we should have 2 containers killable map
    editPolicy.editSchedule();
    Assert.assertEquals(0, editPolicy.getToPreemptContainers().size());
    // Check if previous killable containers included by new killable containers
    Map<ContainerId, RMContainer> killableContainers = waitKillableContainersSize(pm, "a", RMNodeLabelsManager.NO_LABEL, 2);
    Assert.assertTrue(Sets.difference(previousKillableContainers, killableContainers.keySet()).isEmpty());
}

Example 59 with RMNode

use of org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode in project hadoop by apache.

the class TestCapacitySchedulerLazyPreemption method testPreemptionConsidersUserLimit.

@Test(timeout = 60000)
public void testPreemptionConsidersUserLimit() throws Exception {
    /**
     * Test case: Submit two application (app1/app2) to different queues, queue
     * structure:
     *
     * <pre>
     *             Root
     *            /  |  \
     *           a   b   c
     *          10   20  70
     * </pre>
     *
     * Queue-c's user-limit-factor = 0.1, so single user cannot allocate >1 containers in queue-c
     *
     * 1) Two nodes in the cluster, each of them has 4G.
     *
     * 2) app1 submit to queue-a first, it asked 7 * 1G containers, so there's no
     * more resource available.
     *
     * 3) app2 submit to queue-c, ask for one 1G container (for AM)
     *
     * Now the cluster is fulfilled.
     *
     * 4) app2 asks for another 1G container, system will preempt one container
     * from app1, and app2 will receive the preempted container
     */
    CapacitySchedulerConfiguration csConf = new CapacitySchedulerConfiguration(conf);
    csConf.setUserLimitFactor(CapacitySchedulerConfiguration.ROOT + ".c", 0.1f);
    MockRM rm1 = new MockRM(csConf);
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 4 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 4 * 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, 6, new ArrayList<ContainerId>());
    // Do allocation 3 times for node1/node2
    for (int i = 0; i < 3; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
        cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    }
    // App1 should have 7 containers now, and no available resource for cluster
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    Assert.assertEquals(7, schedulerApp1.getLiveContainers().size());
    // Submit app2 to queue-c and asks for a 1G container for AM
    RMApp app2 = rm1.submitApp(1 * GB, "app", "user", null, "c");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm2);
    // NM1/NM2 has available resource = 0G
    Assert.assertEquals(0 * GB, cs.getNode(nm1.getNodeId()).getUnallocatedResource().getMemorySize());
    Assert.assertEquals(0 * GB, cs.getNode(nm2.getNodeId()).getUnallocatedResource().getMemorySize());
    // AM asks for a 1 * GB container
    am2.allocate(Arrays.asList(ResourceRequest.newInstance(Priority.newInstance(1), ResourceRequest.ANY, Resources.createResource(1 * GB), 1)), null);
    // Get edit policy and do one update
    SchedulingEditPolicy editPolicy = getSchedulingEditPolicy(rm1);
    // Call edit schedule twice, and check if no container from app1 marked
    // to be "killable"
    editPolicy.editSchedule();
    editPolicy.editSchedule();
    // No preemption happens
    PreemptionManager pm = cs.getPreemptionManager();
    Map<ContainerId, RMContainer> killableContainers = waitKillableContainersSize(pm, "a", RMNodeLabelsManager.NO_LABEL, 0);
    Assert.assertEquals(0, killableContainers.size());
    // Call CS.handle once to see if container preempted
    cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(am2.getApplicationAttemptId());
    // App1 has 7 containers, and app2 has 1 containers (nothing preempted)
    Assert.assertEquals(7, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(1, schedulerApp2.getLiveContainers().size());
    rm1.close();
}

Example 60 with RMNode

use of org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode in project hadoop by apache.

the class TestCapacitySchedulerSurgicalPreemption method testPriorityPreemptionRequiresMoveReservation.

@Test(timeout = 300000)
public void testPriorityPreemptionRequiresMoveReservation() throws Exception {
    /**
     * Test case: Submit two application (app1/app2) to different queues, queue
     * structure:
     *
     * <pre>
     *             Root
     *            /  |  \
     *           a   b   c
     *          10   20  70
     * </pre>
     *
     * 1) 3 nodes in the cluster, 10G for each
     *
     * 2) app1 submit to queue-b first, it asked 2G each,
     *    it can get 2G on n1 (AM), 2 * 2G on n2
     *
     * 3) app2 submit to queue-c, with 2G AM container (allocated on n3)
     *    app2 requires 9G resource, which will be reserved on n3
     *
     * We should expect container unreserved from n3 and allocated on n1/n2
     */
    conf.setPUOrderingPolicyUnderUtilizedPreemptionEnabled(true);
    conf.setPUOrderingPolicyUnderUtilizedPreemptionDelay(1000);
    conf.setQueueOrderingPolicy(CapacitySchedulerConfiguration.ROOT, CapacitySchedulerConfiguration.QUEUE_PRIORITY_UTILIZATION_ORDERING_POLICY);
    conf.setPUOrderingPolicyUnderUtilizedPreemptionMoveReservation(true);
    // Queue c has higher priority than a/b
    conf.setQueuePriority(CapacitySchedulerConfiguration.ROOT + ".c", 1);
    MockRM rm1 = new MockRM(conf);
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 10 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 10 * GB);
    MockNM nm3 = rm1.registerNode("h3:1234", 10 * GB);
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    RMNode rmNode2 = rm1.getRMContext().getRMNodes().get(nm2.getNodeId());
    RMNode rmNode3 = rm1.getRMContext().getRMNodes().get(nm3.getNodeId());
    // launch an app to queue, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(2 * GB, "app", "user", null, "b");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    am1.allocate("*", 2 * GB, 2, new ArrayList<>());
    // Do allocation for node2 twice
    for (int i = 0; i < 2; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
        cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
    }
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    Assert.assertEquals(3, schedulerApp1.getLiveContainers().size());
    // 1 from n1 and 2 from n2
    waitNumberOfLiveContainersOnNodeFromApp(cs.getNode(rmNode1.getNodeID()), am1.getApplicationAttemptId(), 1);
    waitNumberOfLiveContainersOnNodeFromApp(cs.getNode(rmNode2.getNodeID()), am1.getApplicationAttemptId(), 2);
    // Submit app2 to queue-c and asks for a 2G container for AM, on n3
    RMApp app2 = rm1.submitApp(2 * GB, "app", "user", null, "c");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm3);
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(ApplicationAttemptId.newInstance(app2.getApplicationId(), 1));
    // Asks 1 * 9G container
    am2.allocate("*", 9 * GB, 1, new ArrayList<>());
    // Do allocation for node3 once
    cs.handle(new NodeUpdateSchedulerEvent(rmNode3));
    // Make sure container reserved on node3
    Assert.assertNotNull(cs.getNode(rmNode3.getNodeID()).getReservedContainer());
    // Call editSchedule immediately: nothing happens
    ProportionalCapacityPreemptionPolicy editPolicy = (ProportionalCapacityPreemptionPolicy) getSchedulingEditPolicy(rm1);
    editPolicy.editSchedule();
    Assert.assertNotNull(cs.getNode(rmNode3.getNodeID()).getReservedContainer());
    // Sleep the timeout interval, we should be able to see reserved container
    // moved to n2 (n1 occupied by AM)
    Thread.sleep(1000);
    editPolicy.editSchedule();
    Assert.assertNull(cs.getNode(rmNode3.getNodeID()).getReservedContainer());
    Assert.assertNotNull(cs.getNode(rmNode2.getNodeID()).getReservedContainer());
    Assert.assertEquals(am2.getApplicationAttemptId(), cs.getNode(rmNode2.getNodeID()).getReservedContainer().getApplicationAttemptId());
    // Do it again, we should see containers marked to be preempt
    editPolicy.editSchedule();
    Assert.assertEquals(2, editPolicy.getToPreemptContainers().size());
    // Call editSchedule again: selected containers are killed
    editPolicy.editSchedule();
    // Do allocation till reserved container allocated
    while (schedulerApp2.getLiveContainers().size() < 2) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNode2));
        Thread.sleep(200);
    }
    waitNumberOfLiveContainersFromApp(schedulerApp1, 1);
    rm1.close();
}