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

Example 61 with RMNode

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

the class TestContainerResizing method testSimpleIncreaseContainer.

@Test
public void testSimpleIncreaseContainer() throws Exception {
    /**
     * Application has a container running, and the node has enough available
     * resource. Add a increase request to see if container will be increased
     */
    MockRM rm1 = new MockRM() {

        @Override
        public RMNodeLabelsManager createNodeLabelManager() {
            return mgr;
        }
    };
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 20 * GB);
    // app1 -> a1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    ContainerId containerId1 = ContainerId.newContainerId(am1.getApplicationAttemptId(), 1);
    sentRMContainerLaunched(rm1, containerId1);
    // am1 asks to change its AM container from 1GB to 3GB
    am1.sendContainerResizingRequest(Arrays.asList(UpdateContainerRequest.newInstance(0, containerId1, ContainerUpdateType.INCREASE_RESOURCE, Resources.createResource(3 * GB), null)));
    FiCaSchedulerApp app = TestUtils.getFiCaSchedulerApp(rm1, app1.getApplicationId());
    checkPendingResource(rm1, "default", 2 * GB, null);
    Assert.assertEquals(2 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // NM1 do 1 heartbeats
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    // Pending resource should be deducted
    checkPendingResource(rm1, "default", 0 * GB, null);
    Assert.assertEquals(0 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    verifyContainerIncreased(am1.allocate(null, null), containerId1, 3 * GB);
    verifyAvailableResourceOfSchedulerNode(rm1, nm1.getNodeId(), 17 * GB);
    rm1.close();
}

Example 62 with RMNode

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

the class TestContainerResizing method testExcessiveReservationWhenCancelIncreaseRequest.

@Test
public void testExcessiveReservationWhenCancelIncreaseRequest() throws Exception {
    /**
     * Application has two containers running, try to increase one of then, node
     * doesn't have enough resource, so the increase request will be reserved.
     * Check resource usage after container reserved, finish a container &
     * cancel the increase request, reservation should be cancelled
     */
    MockRM rm1 = new MockRM() {

        @Override
        public RMNodeLabelsManager createNodeLabelManager() {
            return mgr;
        }
    };
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 8 * GB);
    // app1 -> a1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    FiCaSchedulerApp app = TestUtils.getFiCaSchedulerApp(rm1, app1.getApplicationId());
    // Allocate two more containers
    am1.allocate(Arrays.asList(ResourceRequest.newInstance(Priority.newInstance(1), "*", Resources.createResource(2 * GB), 1)), null);
    ContainerId containerId2 = ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
    Assert.assertTrue(rm1.waitForState(nm1, containerId2, RMContainerState.ALLOCATED));
    // Acquire them, and NM report RUNNING
    am1.allocate(null, null);
    sentRMContainerLaunched(rm1, containerId2);
    ContainerId containerId1 = ContainerId.newContainerId(am1.getApplicationAttemptId(), 1);
    sentRMContainerLaunched(rm1, containerId1);
    // am1 asks to change its AM container from 1GB to 7GB
    am1.sendContainerResizingRequest(Arrays.asList(UpdateContainerRequest.newInstance(0, containerId1, ContainerUpdateType.INCREASE_RESOURCE, Resources.createResource(7 * GB), null)));
    checkPendingResource(rm1, "default", 6 * GB, null);
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // NM1 do 1 heartbeats
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    /* Check reservation statuses */
    // Increase request should be reserved
    Assert.assertFalse(app.getReservedContainers().isEmpty());
    Assert.assertNotNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Pending resource will not be changed since it's not satisfied
    checkPendingResource(rm1, "default", 6 * GB, null);
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // Queue/user/application's usage will be updated
    checkUsedResource(rm1, "default", 9 * GB, null);
    Assert.assertEquals(9 * GB, ((LeafQueue) cs.getQueue("default")).getUser("user").getUsed().getMemorySize());
    Assert.assertEquals(3 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getReserved().getMemorySize());
    // Complete one container and cancel increase request (via send a increase
    // request, make target_capacity=existing_capacity)
    am1.allocate(null, Arrays.asList(containerId2));
    // am1 asks to change its AM container from 1G to 1G (cancel the increase
    // request actually)
    am1.sendContainerResizingRequest(Arrays.asList(UpdateContainerRequest.newInstance(0, containerId1, ContainerUpdateType.INCREASE_RESOURCE, Resources.createResource(1 * GB), null)));
    // Trigger a node heartbeat..
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    /* Check statuses after reservation satisfied */
    // Increase request should be unreserved
    Assert.assertTrue(app.getReservedContainers().isEmpty());
    Assert.assertNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Pending resource will be changed since it's satisfied
    checkPendingResource(rm1, "default", 0 * GB, null);
    Assert.assertEquals(0 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // Queue/user/application's usage will be updated
    checkUsedResource(rm1, "default", 1 * GB, null);
    Assert.assertEquals(1 * GB, ((LeafQueue) cs.getQueue("default")).getUser("user").getUsed().getMemorySize());
    Assert.assertEquals(0 * GB, app.getAppAttemptResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(1 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    rm1.close();
}

Example 63 with RMNode

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

the class TestContainerResizing method testIncreaseContainerUnreservedWhenContainerCompleted.

@Test
public void testIncreaseContainerUnreservedWhenContainerCompleted() throws Exception {
    /**
     * App has two containers on the same node (node.resource = 8G), container1
     * = 2G, container2 = 2G. App asks to increase container2 to 8G.
     *
     * So increase container request will be reserved. When app releases
     * container2, reserved part should be released as well.
     */
    MockRM rm1 = new MockRM() {

        @Override
        public RMNodeLabelsManager createNodeLabelManager() {
            return mgr;
        }
    };
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
    MockNM nm2 = rm1.registerNode("h2:1234", 8 * GB);
    // app1 -> a1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    FiCaSchedulerApp app = TestUtils.getFiCaSchedulerApp(rm1, app1.getApplicationId());
    // Allocate two more containers
    am1.allocate(Arrays.asList(ResourceRequest.newInstance(Priority.newInstance(1), "*", Resources.createResource(2 * GB), 1)), null);
    ContainerId containerId2 = ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
    Assert.assertTrue(rm1.waitForState(nm1, containerId2, RMContainerState.ALLOCATED));
    // Acquire them, and NM report RUNNING
    am1.allocate(null, null);
    sentRMContainerLaunched(rm1, containerId2);
    rm1.waitForState(Arrays.asList(nm1, nm2), containerId2, RMContainerState.RUNNING);
    // am1 asks to change its AM container from 2GB to 8GB
    am1.sendContainerResizingRequest(Arrays.asList(UpdateContainerRequest.newInstance(0, containerId2, ContainerUpdateType.INCREASE_RESOURCE, Resources.createResource(8 * GB), null)));
    checkPendingResource(rm1, "default", 6 * GB, null);
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // NM1 do 1 heartbeats
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    /* Check reservation statuses */
    // Increase request should be reserved
    Assert.assertFalse(app.getReservedContainers().isEmpty());
    Assert.assertNotNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Pending resource will not be changed since it's not satisfied
    checkPendingResource(rm1, "default", 6 * GB, null);
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // Queue/user/application's usage will be updated
    checkUsedResource(rm1, "default", 9 * GB, null);
    Assert.assertEquals(9 * GB, ((LeafQueue) cs.getQueue("default")).getUser("user").getUsed().getMemorySize());
    Assert.assertEquals(3 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getReserved().getMemorySize());
    // Complete container2, container will be unreserved and completed
    am1.allocate(null, Arrays.asList(containerId2));
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    am1.allocate(null, null);
    /* Check statuses after reservation satisfied */
    // Increase request should be unreserved
    Assert.assertTrue(app.getReservedContainers().isEmpty());
    Assert.assertNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Pending resource will be changed since it's satisfied
    checkPendingResource(rm1, "default", 0 * GB, null);
    Assert.assertEquals(0 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // Queue/user/application's usage will be updated
    checkUsedResource(rm1, "default", 1 * GB, null);
    Assert.assertEquals(1 * GB, ((LeafQueue) cs.getQueue("default")).getUser("user").getUsed().getMemorySize());
    Assert.assertEquals(0 * GB, app.getAppAttemptResourceUsage().getReserved().getMemorySize());
    Assert.assertEquals(1 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    rm1.close();
}

Example 64 with RMNode

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

the class TestContainerResizing method testIncreaseRequestWithNoHeadroomLeft.

@Test
public void testIncreaseRequestWithNoHeadroomLeft() throws Exception {
    /**
     * Application has two containers running, try to increase one of them, the
     * requested amount exceeds user's headroom for the queue.
     */
    MockRM rm1 = new MockRM() {

        @Override
        public RMNodeLabelsManager createNodeLabelManager() {
            return mgr;
        }
    };
    rm1.start();
    MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
    // app1 -> a1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    FiCaSchedulerApp app = TestUtils.getFiCaSchedulerApp(rm1, app1.getApplicationId());
    // Allocate 1 container
    am1.allocate(Arrays.asList(ResourceRequest.newInstance(Priority.newInstance(1), "*", Resources.createResource(2 * GB), 1)), null);
    ContainerId containerId2 = ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
    Assert.assertTrue(rm1.waitForState(nm1, containerId2, RMContainerState.ALLOCATED));
    // Acquire them, and NM report RUNNING
    am1.allocate(null, null);
    sentRMContainerLaunched(rm1, containerId2);
    // am1 asks to change container2 from 2GB to 8GB, which will exceed user
    // limit
    am1.sendContainerResizingRequest(Arrays.asList(UpdateContainerRequest.newInstance(0, containerId2, ContainerUpdateType.INCREASE_RESOURCE, Resources.createResource(8 * GB), null)));
    checkPendingResource(rm1, "default", 6 * GB, null);
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // NM1 do 1 heartbeats
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    /* Check reservation statuses */
    // Increase request should *NOT* be reserved as it exceeds user limit
    Assert.assertTrue(app.getReservedContainers().isEmpty());
    Assert.assertNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
    // Pending resource will not be changed since it's not satisfied
    checkPendingResource(rm1, "default", 6 * GB, null);
    Assert.assertEquals(6 * GB, app.getAppAttemptResourceUsage().getPending().getMemorySize());
    // Queue/user/application's usage will *NOT* be updated
    checkUsedResource(rm1, "default", 3 * GB, null);
    Assert.assertEquals(3 * GB, ((LeafQueue) cs.getQueue("default")).getUser("user").getUsed().getMemorySize());
    Assert.assertEquals(3 * GB, app.getAppAttemptResourceUsage().getUsed().getMemorySize());
    Assert.assertEquals(0 * GB, app.getAppAttemptResourceUsage().getReserved().getMemorySize());
    rm1.close();
}

Example 65 with RMNode

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

the class TestNodeLabelContainerAllocation method testContainerReservationWithLabels.

@Test(timeout = 120000)
public void testContainerReservationWithLabels() throws Exception {
    // This test is pretty much similar to testContainerAllocateWithLabel.
    // Difference is, this test doesn't specify label expression in
    // ResourceRequest,
    // instead, it uses default queue label expression
    // set node -> label
    mgr.addToCluserNodeLabelsWithDefaultExclusivity(ImmutableSet.of("x", "y", "z"));
    mgr.addLabelsToNode(ImmutableMap.of(NodeId.newInstance("h1", 0), toSet("x"), NodeId.newInstance("h2", 0), toSet("y"), NodeId.newInstance("h3", 0), toSet("x")));
    // inject node label manager
    MockRM rm1 = new MockRM(TestUtils.getConfigurationWithDefaultQueueLabels(conf)) {

        @Override
        public RMNodeLabelsManager createNodeLabelManager() {
            return mgr;
        }
    };
    rm1.getRMContext().setNodeLabelManager(mgr);
    rm1.start();
    // label = x
    MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
    // label = y
    rm1.registerNode("h2:1234", 8 * GB);
    // label = x
    rm1.registerNode("h3:1234", 8 * GB);
    ContainerId containerId;
    // launch an app to queue a1 (label = x), and check all container will
    // be allocated in h1
    RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "a1");
    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
    // request a container.
    am1.allocate("*", 4 * GB, 2, new ArrayList<ContainerId>());
    containerId = ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
    LeafQueue leafQueue = (LeafQueue) cs.getQueue("a1");
    // Do node heartbeats 2 times
    // First time will allocate container for app1, second time will reserve
    // container for app1
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    checkTaskContainersHost(am1.getApplicationAttemptId(), containerId, rm1, "h1");
    // Check if a 4G container allocated for app1, and 4G is reserved
    FiCaSchedulerApp schedulerApp1 = cs.getApplicationAttempt(am1.getApplicationAttemptId());
    Assert.assertEquals(2, schedulerApp1.getLiveContainers().size());
    Assert.assertTrue(schedulerApp1.getReservedContainers().size() > 0);
    Assert.assertEquals(9 * GB, cs.getRootQueue().getQueueResourceUsage().getUsed("x").getMemorySize());
    Assert.assertEquals(4 * GB, cs.getRootQueue().getQueueResourceUsage().getReserved("x").getMemorySize());
    Assert.assertEquals(4 * GB, leafQueue.getQueueResourceUsage().getReserved("x").getMemorySize());
    // Cancel asks of app2 and re-kick RM
    am1.allocate("*", 4 * GB, 0, new ArrayList<ContainerId>());
    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
    Assert.assertEquals(5 * GB, cs.getRootQueue().getQueueResourceUsage().getUsed("x").getMemorySize());
    Assert.assertEquals(0, cs.getRootQueue().getQueueResourceUsage().getReserved("x").getMemorySize());
    Assert.assertEquals(0, leafQueue.getQueueResourceUsage().getReserved("x").getMemorySize());
    rm1.close();
}