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

Example 86 with MockAM

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

the class TestCapacitySchedulerSurgicalPreemption method testPriorityPreemptionFromHighestPriorityQueueAndOldestContainer.

@Test(timeout = 600000)
public void testPriorityPreemptionFromHighestPriorityQueueAndOldestContainer() throws Exception {
    /**
     * Test case: Submit two application (app1/app2) to different queues, queue
     * structure:
     *
     * <pre>
     *             Root
     *            /  |  \
     *           a   b   c
     *          45  45  10
     * </pre>
     *
     * Priority of queue_a = 1
     * Priority of queue_b = 2
     *
     * 1) 5 nodes (n0-n4) in the cluster, each of them has 4G.
     *
     * 2) app1 submit to queue-c first (AM=1G), it asked 4 * 1G containers
     *    We will allocate 1 container on each of n0-n4. AM on n4.
     *
     * 3) app2 submit to queue-a, AM container=0.5G, allocated on n0
     *    Ask for 2 * 3.5G containers. (Reserved on n0/n1)
     *
     * 4) app2 submit to queue-b, AM container=0.5G, allocated on n2
     *    Ask for 2 * 3.5G containers. (Reserved on n2/n3)
     *
     * First we will preempt container on n2 since it is the oldest container of
     * Highest priority queue (b)
     */
    // Total preemption = 1G per round, which is 5% of cluster resource (20G)
    conf.setFloat(CapacitySchedulerConfiguration.TOTAL_PREEMPTION_PER_ROUND, 0.05f);
    conf.setInt(YarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_MB, 512);
    conf.setPUOrderingPolicyUnderUtilizedPreemptionEnabled(true);
    conf.setPUOrderingPolicyUnderUtilizedPreemptionDelay(1000);
    conf.setQueueOrderingPolicy(CapacitySchedulerConfiguration.ROOT, CapacitySchedulerConfiguration.QUEUE_PRIORITY_UTILIZATION_ORDERING_POLICY);
    // A/B has higher priority
    conf.setQueuePriority(CapacitySchedulerConfiguration.ROOT + ".a", 1);
    conf.setQueuePriority(CapacitySchedulerConfiguration.ROOT + ".b", 2);
    conf.setCapacity(CapacitySchedulerConfiguration.ROOT + ".a", 45f);
    conf.setCapacity(CapacitySchedulerConfiguration.ROOT + ".b", 45f);
    conf.setCapacity(CapacitySchedulerConfiguration.ROOT + ".c", 10f);
    MockRM rm1 = new MockRM(conf);
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    MockNM[] mockNMs = new MockNM[5];
    for (int i = 0; i < 5; i++) {
        mockNMs[i] = rm1.registerNode("h" + i + ":1234", 4 * GB);
    }
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode[] rmNodes = new RMNode[5];
    for (int i = 0; i < 5; i++) {
        rmNodes[i] = rm1.getRMContext().getRMNodes().get(mockNMs[i].getNodeId());
    }
    // launch an app to queue, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "c");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, mockNMs[4]);
    am1.allocate("*", 1 * GB, 4, new ArrayList<>());
    // Do allocation for nm1-nm8
    for (int i = 0; i < 4; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNodes[i]));
    }
    // App1 should have 5 containers now, one for each node
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    Assert.assertEquals(5, schedulerApp1.getLiveContainers().size());
    for (int i = 0; i < 5; i++) {
        waitNumberOfLiveContainersOnNodeFromApp(cs.getNode(rmNodes[i].getNodeID()), am1.getApplicationAttemptId(), 1);
    }
    // Submit app2 to queue-a and asks for a 0.5G container for AM (on n0)
    RMApp app2 = rm1.submitApp(512, "app", "user", null, "a");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, mockNMs[0]);
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(ApplicationAttemptId.newInstance(app2.getApplicationId(), 1));
    // Ask 2 * 3.5GB containers
    am2.allocate("*", 3 * GB + 512, 2, new ArrayList<>());
    // Do allocation for n0-n1
    for (int i = 0; i < 2; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNodes[i]));
    }
    // Check am2 reserved resource from nm0-nm1
    for (int i = 0; i < 2; i++) {
        Assert.assertNotNull("Should reserve on nm-" + i, cs.getNode(rmNodes[i].getNodeID()).getReservedContainer());
        Assert.assertEquals(cs.getNode(rmNodes[i].getNodeID()).getReservedContainer().getQueueName(), "a");
    }
    // Submit app3 to queue-b and asks for a 0.5G container for AM (on n2)
    RMApp app3 = rm1.submitApp(512, "app", "user", null, "b");
    MockAM am3 = MockRM.launchAndRegisterAM(app3, rm1, mockNMs[2]);
    FiCaSchedulerApp schedulerApp3 = cs.getApplicationAttempt(ApplicationAttemptId.newInstance(app3.getApplicationId(), 1));
    // Ask 2 * 3.5GB containers
    am3.allocate("*", 3 * GB + 512, 2, new ArrayList<>());
    // Do allocation for n2-n3
    for (int i = 2; i < 4; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNodes[i]));
    }
    // Check am2 reserved resource from nm2-nm3
    for (int i = 2; i < 4; i++) {
        Assert.assertNotNull("Should reserve on nm-" + i, cs.getNode(rmNodes[i].getNodeID()).getReservedContainer());
        Assert.assertEquals(cs.getNode(rmNodes[i].getNodeID()).getReservedContainer().getQueueName(), "b");
    }
    // Sleep the timeout interval, we should be able to see 1 container selected
    Thread.sleep(1000);
    /* 1st container preempted is on n2 */
    ProportionalCapacityPreemptionPolicy editPolicy = (ProportionalCapacityPreemptionPolicy) getSchedulingEditPolicy(rm1);
    editPolicy.editSchedule();
    // We should have one to-preempt container, on node[2]
    Set<RMContainer> selectedToPreempt = editPolicy.getToPreemptContainers().keySet();
    Assert.assertEquals(1, selectedToPreempt.size());
    Assert.assertEquals(mockNMs[2].getNodeId(), selectedToPreempt.iterator().next().getAllocatedNode());
    // Call editSchedule again: selected containers are killed
    editPolicy.editSchedule();
    waitNumberOfLiveContainersFromApp(schedulerApp1, 4);
    // Make sure the container killed, then do allocation for all nms
    for (int i = 0; i < 4; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNodes[i]));
    }
    waitNumberOfLiveContainersFromApp(schedulerApp1, 4);
    waitNumberOfLiveContainersFromApp(schedulerApp2, 1);
    waitNumberOfLiveContainersFromApp(schedulerApp3, 2);
    /* 2nd container preempted is on n3 */
    editPolicy.editSchedule();
    // We should have one to-preempt container, on node[3]
    selectedToPreempt = editPolicy.getToPreemptContainers().keySet();
    Assert.assertEquals(1, selectedToPreempt.size());
    Assert.assertEquals(mockNMs[3].getNodeId(), selectedToPreempt.iterator().next().getAllocatedNode());
    // Call editSchedule again: selected containers are killed
    editPolicy.editSchedule();
    waitNumberOfLiveContainersFromApp(schedulerApp1, 3);
    // Do allocation for all nms
    for (int i = 0; i < 4; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNodes[i]));
    }
    waitNumberOfLiveContainersFromApp(schedulerApp1, 3);
    waitNumberOfLiveContainersFromApp(schedulerApp2, 1);
    waitNumberOfLiveContainersFromApp(schedulerApp3, 3);
    /* 3rd container preempted is on n0 */
    editPolicy.editSchedule();
    // We should have one to-preempt container, on node[0]
    selectedToPreempt = editPolicy.getToPreemptContainers().keySet();
    Assert.assertEquals(1, selectedToPreempt.size());
    Assert.assertEquals(mockNMs[0].getNodeId(), selectedToPreempt.iterator().next().getAllocatedNode());
    // Call editSchedule again: selected containers are killed
    editPolicy.editSchedule();
    waitNumberOfLiveContainersFromApp(schedulerApp1, 2);
    // Do allocation for all nms
    for (int i = 0; i < 4; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNodes[i]));
    }
    waitNumberOfLiveContainersFromApp(schedulerApp1, 2);
    waitNumberOfLiveContainersFromApp(schedulerApp2, 2);
    waitNumberOfLiveContainersFromApp(schedulerApp3, 3);
    /* 4th container preempted is on n1 */
    editPolicy.editSchedule();
    // We should have one to-preempt container, on node[0]
    selectedToPreempt = editPolicy.getToPreemptContainers().keySet();
    Assert.assertEquals(1, selectedToPreempt.size());
    Assert.assertEquals(mockNMs[1].getNodeId(), selectedToPreempt.iterator().next().getAllocatedNode());
    // Call editSchedule again: selected containers are killed
    editPolicy.editSchedule();
    waitNumberOfLiveContainersFromApp(schedulerApp1, 1);
    // Do allocation for all nms
    for (int i = 0; i < 4; i++) {
        cs.handle(new NodeUpdateSchedulerEvent(rmNodes[i]));
    }
    waitNumberOfLiveContainersFromApp(schedulerApp1, 1);
    waitNumberOfLiveContainersFromApp(schedulerApp2, 3);
    waitNumberOfLiveContainersFromApp(schedulerApp3, 3);
    rm1.close();
}

Example 87 with MockAM

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

the class TestContainerAllocation method testPendingResourcesConsideringUserLimit.

@Test
public void testPendingResourcesConsideringUserLimit() throws Exception {
    // Set maximum capacity of A to 10
    CapacitySchedulerConfiguration newConf = new CapacitySchedulerConfiguration(conf);
    newConf.setUserLimitFactor(CapacitySchedulerConfiguration.ROOT + ".default", 0.5f);
    newConf.setMaximumAMResourcePercentPerPartition(CapacitySchedulerConfiguration.ROOT + ".default", "", 1.0f);
    MockRM rm1 = new MockRM(newConf);
    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 default, AM container should be launched in nm1
    RMApp app1 = rm1.submitApp(2 * GB, "app", "u1", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    // launch 2nd app to queue default, AM container should be launched in nm1
    RMApp app2 = rm1.submitApp(4 * GB, "app", "u2", null, "default");
    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm1);
    // am1 asks 1 * 3G container
    am1.allocate("*", 3 * GB, 1, null);
    // am2 asks 4 * 5G container
    am2.allocate("*", 5 * GB, 4, null);
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    RMNode rmNode2 = rm1.getRMContext().getRMNodes().get(nm2.getNodeId());
    // Do node heartbeats one, we expect one container allocated reserved on nm1
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    // App1 will get 1 container reserved
    Assert.assertEquals(1, schedulerApp1.getReservedContainers().size());
    /*
     * Note that the behavior of appAttemptResourceUsage is different from queue's
     * For queue, used = actual-used + reserved
     * For app, used = actual-used.
     *
     * TODO (wangda): Need to make behaviors of queue/app's resource usage
     * consistent
     */
    Assert.assertEquals(2 * GB, schedulerApp1.getAppAttemptResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(3 * GB, schedulerApp1.getAppAttemptResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(3 * GB, schedulerApp1.getAppAttemptResourceUsage().getPending().getMemorySize());
    FiCaSchedulerApp schedulerApp2 = cs.getApplicationAttempt(am2.getApplicationAttemptId());
    Assert.assertEquals(4 * GB, schedulerApp2.getAppAttemptResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(0 * GB, schedulerApp2.getAppAttemptResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(5 * 4 * GB, schedulerApp2.getAppAttemptResourceUsage().getPending().getMemorySize());
    LeafQueue lq = (LeafQueue) cs.getQueue("default");
    // UL = 8GB, so head room of u1 = 8GB - 2GB (AM) - 3GB (Reserved) = 3GB
    //                           u2 = 8GB - 4GB = 4GB
    // When not deduct reserved, total-pending = 3G (u1) + 4G (u2) = 7G
    //          deduct reserved, total-pending = 0G (u1) + 4G = 4G
    Assert.assertEquals(7 * GB, lq.getTotalPendingResourcesConsideringUserLimit(Resources.createResource(20 * GB), "", false).getMemorySize());
    Assert.assertEquals(4 * GB, lq.getTotalPendingResourcesConsideringUserLimit(Resources.createResource(20 * GB), "", true).getMemorySize());
    rm1.close();
}

Example 88 with MockAM

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

the class TestContainerAllocation method testNormalContainerAllocationWhenDNSUnavailable.

@Test
public void testNormalContainerAllocationWhenDNSUnavailable() throws Exception {
    MockRM rm1 = new MockRM(conf);
    rm1.start();
    MockNM nm1 = rm1.registerNode("unknownhost:1234", 8000);
    RMApp app1 = rm1.submitApp(200);
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    // request a container.
    am1.allocate("127.0.0.1", 1024, 1, new ArrayList<ContainerId>());
    ContainerId containerId2 = ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
    rm1.waitForState(nm1, containerId2, RMContainerState.ALLOCATED);
    // acquire the container.
    SecurityUtilTestHelper.setTokenServiceUseIp(true);
    List<Container> containers;
    try {
        containers = am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>()).getAllocatedContainers();
        // not able to fetch the container;
        Assert.assertEquals(0, containers.size());
    } finally {
        SecurityUtilTestHelper.setTokenServiceUseIp(false);
    }
    containers = am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>()).getAllocatedContainers();
    // should be able to fetch the container;
    Assert.assertEquals(1, containers.size());
}

Example 89 with MockAM

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

the class TestContainerAllocation method testExcessReservationThanNodeManagerCapacity.

@Test(timeout = 60000)
public void testExcessReservationThanNodeManagerCapacity() throws Exception {
    @SuppressWarnings("resource") MockRM rm = new MockRM(conf);
    rm.start();
    // Register node1
    MockNM nm1 = rm.registerNode("127.0.0.1:1234", 2 * GB, 4);
    MockNM nm2 = rm.registerNode("127.0.0.1:2234", 3 * GB, 4);
    nm1.nodeHeartbeat(true);
    nm2.nodeHeartbeat(true);
    // wait..
    int waitCount = 20;
    int size = rm.getRMContext().getRMNodes().size();
    while ((size = rm.getRMContext().getRMNodes().size()) != 2 && waitCount-- > 0) {
        LOG.info("Waiting for node managers to register : " + size);
        Thread.sleep(100);
    }
    Assert.assertEquals(2, rm.getRMContext().getRMNodes().size());
    // Submit an application
    RMApp app1 = rm.submitApp(128);
    // kick the scheduling
    nm1.nodeHeartbeat(true);
    RMAppAttempt attempt1 = app1.getCurrentAppAttempt();
    MockAM am1 = rm.sendAMLaunched(attempt1.getAppAttemptId());
    am1.registerAppAttempt();
    LOG.info("sending container requests ");
    am1.addRequests(new String[] { "*" }, 2 * GB, 1, 1);
    // send the request
    AllocateResponse alloc1Response = am1.schedule();
    // kick the scheduler
    nm1.nodeHeartbeat(true);
    int waitCounter = 20;
    LOG.info("heartbeating nm1");
    while (alloc1Response.getAllocatedContainers().size() < 1 && waitCounter-- > 0) {
        LOG.info("Waiting for containers to be created for app 1...");
        Thread.sleep(500);
        alloc1Response = am1.schedule();
    }
    LOG.info("received container : " + alloc1Response.getAllocatedContainers().size());
    // No container should be allocated.
    // Internally it should not been reserved.
    Assert.assertTrue(alloc1Response.getAllocatedContainers().size() == 0);
    LOG.info("heartbeating nm2");
    waitCounter = 20;
    nm2.nodeHeartbeat(true);
    while (alloc1Response.getAllocatedContainers().size() < 1 && waitCounter-- > 0) {
        LOG.info("Waiting for containers to be created for app 1...");
        Thread.sleep(500);
        alloc1Response = am1.schedule();
    }
    LOG.info("received container : " + alloc1Response.getAllocatedContainers().size());
    Assert.assertTrue(alloc1Response.getAllocatedContainers().size() == 1);
    rm.stop();
}

Example 90 with MockAM

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

the class TestContainerAllocation method testAllocationForReservedContainer.

@Test(timeout = 60000)
public void testAllocationForReservedContainer() 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());
    // Mark one app1 container as killed/completed and re-kick RM
    for (RMContainer container : schedulerApp1.getLiveContainers()) {
        if (container.isAMContainer()) {
            continue;
        }
        cs.markContainerForKillable(container);
    }
    // Cancel asks of app1 and re-kick RM
    am1.allocate("*", 4 * GB, 0, new ArrayList<ContainerId>());
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    // Check 4G container cancelled for app1, and one container allocated for
    // app2
    Assert.assertEquals(1, schedulerApp1.getLiveContainers().size());
    Assert.assertEquals(2, schedulerApp2.getLiveContainers().size());
    Assert.assertFalse(schedulerApp2.getReservedContainers().size() > 0);
    // NM1 has available resource = 2G (8G - 2 * 1G - 4G)
    Assert.assertEquals(2 * GB, cs.getNode(nm1.getNodeId()).getUnallocatedResource().getMemorySize());
    Assert.assertNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Usage of queue = 4G + 2 * 1G
    Assert.assertEquals(6 * 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();
}