Search in sources :

Example 51 with ApplicationAttemptId

use of org.apache.hadoop.yarn.api.records.ApplicationAttemptId 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 52 with ApplicationAttemptId

use of org.apache.hadoop.yarn.api.records.ApplicationAttemptId 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)

Example 53 with ApplicationAttemptId

use of org.apache.hadoop.yarn.api.records.ApplicationAttemptId in project hadoop by apache.

the class TestLeafQueue method testSingleQueueWithMultipleUsers.

@Test
public void testSingleQueueWithMultipleUsers() throws Exception {
    // Mock the queue
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(A));
    //unset maxCapacity
    a.setMaxCapacity(1.0f);
    // Users
    final String user_0 = "user_0";
    final String user_1 = "user_1";
    final String user_2 = "user_2";
    // Submit applications
    final ApplicationAttemptId appAttemptId_0 = TestUtils.getMockApplicationAttemptId(0, 0);
    FiCaSchedulerApp app_0 = new FiCaSchedulerApp(appAttemptId_0, user_0, a, a.getAbstractUsersManager(), 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, a.getAbstractUsersManager(), spyRMContext);
    // same user
    a.submitApplicationAttempt(app_1, user_0);
    final ApplicationAttemptId appAttemptId_2 = TestUtils.getMockApplicationAttemptId(2, 0);
    FiCaSchedulerApp app_2 = new FiCaSchedulerApp(appAttemptId_2, user_1, a, a.getAbstractUsersManager(), spyRMContext);
    a.submitApplicationAttempt(app_2, user_1);
    final ApplicationAttemptId appAttemptId_3 = TestUtils.getMockApplicationAttemptId(3, 0);
    FiCaSchedulerApp app_3 = new FiCaSchedulerApp(appAttemptId_3, user_2, a, a.getAbstractUsersManager(), spyRMContext);
    a.submitApplicationAttempt(app_3, user_2);
    Map<ApplicationAttemptId, FiCaSchedulerApp> apps = ImmutableMap.of(app_0.getApplicationAttemptId(), app_0, app_1.getApplicationAttemptId(), app_1, app_2.getApplicationAttemptId(), app_2, app_3.getApplicationAttemptId(), app_3);
    // Setup some nodes
    String host_0 = "127.0.0.1";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, DEFAULT_RACK, 0, 8 * GB);
    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);
    when(csContext.getClusterResource()).thenReturn(clusterResource);
    // Setup resource-requests
    Priority priority = TestUtils.createMockPriority(1);
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 10, true, priority, recordFactory)));
    app_1.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 10, 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(1 * GB, a.getUsedResources().getMemorySize());
    assertEquals(1 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    // Also 2nd -> minCapacity = 1024 since (.1 * 8G) < minAlloc, also
    // you can get one container more than user-limit
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(2 * GB, a.getUsedResources().getMemorySize());
    assertEquals(2 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    // Can't allocate 3rd due to user-limit
    a.setUserLimit(25);
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(2 * GB, a.getUsedResources().getMemorySize());
    assertEquals(2 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    // Submit resource requests for other apps now to 'activate' them
    app_2.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 3 * GB, 1, true, priority, recordFactory)));
    app_3.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority, recordFactory)));
    // Now allocations should goto app_2 since 
    // user_0 is at limit inspite of high user-limit-factor
    a.setUserLimitFactor(10);
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(5 * GB, a.getUsedResources().getMemorySize());
    assertEquals(2 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(3 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_3.getCurrentConsumption().getMemorySize());
    // Now allocations should goto app_0 since 
    // user_0 is at user-limit not above it
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(6 * GB, a.getUsedResources().getMemorySize());
    assertEquals(3 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(3 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_3.getCurrentConsumption().getMemorySize());
    // Test max-capacity
    // Now - no more allocs since we are at max-cap
    a.setMaxCapacity(0.5f);
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(6 * GB, a.getUsedResources().getMemorySize());
    assertEquals(3 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(3 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_3.getCurrentConsumption().getMemorySize());
    // Revert max-capacity and user-limit-factor
    // Now, allocations should goto app_3 since it's under user-limit 
    a.setMaxCapacity(1.0f);
    a.setUserLimitFactor(1);
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(7 * GB, a.getUsedResources().getMemorySize());
    assertEquals(3 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(3 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(1 * GB, app_3.getCurrentConsumption().getMemorySize());
    // Now we should assign to app_3 again since user_2 is under user-limit
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(8 * GB, a.getUsedResources().getMemorySize());
    assertEquals(3 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(3 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(2 * GB, app_3.getCurrentConsumption().getMemorySize());
    // 8. Release each container from app_0
    for (RMContainer rmContainer : app_0.getLiveContainers()) {
        a.completedContainer(clusterResource, app_0, node_0, rmContainer, ContainerStatus.newInstance(rmContainer.getContainerId(), ContainerState.COMPLETE, "", ContainerExitStatus.KILLED_BY_RESOURCEMANAGER), RMContainerEventType.KILL, null, true);
    }
    assertEquals(5 * GB, a.getUsedResources().getMemorySize());
    assertEquals(0 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(3 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(2 * GB, app_3.getCurrentConsumption().getMemorySize());
    // 9. Release each container from app_2
    for (RMContainer rmContainer : app_2.getLiveContainers()) {
        a.completedContainer(clusterResource, app_2, node_0, rmContainer, ContainerStatus.newInstance(rmContainer.getContainerId(), ContainerState.COMPLETE, "", ContainerExitStatus.KILLED_BY_RESOURCEMANAGER), RMContainerEventType.KILL, null, true);
    }
    assertEquals(2 * GB, a.getUsedResources().getMemorySize());
    assertEquals(0 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(2 * GB, app_3.getCurrentConsumption().getMemorySize());
    // 10. Release each container from app_3
    for (RMContainer rmContainer : app_3.getLiveContainers()) {
        a.completedContainer(clusterResource, app_3, node_0, rmContainer, ContainerStatus.newInstance(rmContainer.getContainerId(), ContainerState.COMPLETE, "", ContainerExitStatus.KILLED_BY_RESOURCEMANAGER), RMContainerEventType.KILL, null, true);
    }
    assertEquals(0 * GB, a.getUsedResources().getMemorySize());
    assertEquals(0 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_2.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_3.getCurrentConsumption().getMemorySize());
}
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) RMContainer(org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainer) 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) Test(org.junit.Test)

Example 54 with ApplicationAttemptId

use of org.apache.hadoop.yarn.api.records.ApplicationAttemptId in project hadoop by apache.

the class TestLeafQueue method createListOfApps.

private List<FiCaSchedulerApp> createListOfApps(int noOfApps, String user, LeafQueue defaultQueue) {
    List<FiCaSchedulerApp> appsLists = new ArrayList<FiCaSchedulerApp>();
    for (int i = 0; i < noOfApps; i++) {
        ApplicationAttemptId appAttemptId_0 = TestUtils.getMockApplicationAttemptId(i, 0);
        FiCaSchedulerApp app_0 = new FiCaSchedulerApp(appAttemptId_0, user, defaultQueue, mock(ActiveUsersManager.class), spyRMContext);
        appsLists.add(app_0);
    }
    return appsLists;
}
Also used : FiCaSchedulerApp(org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp) ArrayList(java.util.ArrayList) ActiveUsersManager(org.apache.hadoop.yarn.server.resourcemanager.scheduler.ActiveUsersManager) ApplicationAttemptId(org.apache.hadoop.yarn.api.records.ApplicationAttemptId)

Example 55 with ApplicationAttemptId

use of org.apache.hadoop.yarn.api.records.ApplicationAttemptId in project hadoop by apache.

the class TestLeafQueue method testUserHeadroomMultiApp.

@Test
public void testUserHeadroomMultiApp() throws Exception {
    // Mock the queue
    LeafQueue a = stubLeafQueue((LeafQueue) queues.get(A));
    //unset maxCapacity
    a.setMaxCapacity(1.0f);
    // Users
    final String user_0 = "user_0";
    final String user_1 = "user_1";
    // Submit applications
    final ApplicationAttemptId appAttemptId_0 = TestUtils.getMockApplicationAttemptId(0, 0);
    FiCaSchedulerApp app_0 = new FiCaSchedulerApp(appAttemptId_0, user_0, a, a.getAbstractUsersManager(), 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, a.getAbstractUsersManager(), spyRMContext);
    // same user
    a.submitApplicationAttempt(app_1, user_0);
    final ApplicationAttemptId appAttemptId_2 = TestUtils.getMockApplicationAttemptId(2, 0);
    FiCaSchedulerApp app_2 = new FiCaSchedulerApp(appAttemptId_2, user_1, a, a.getAbstractUsersManager(), spyRMContext);
    a.submitApplicationAttempt(app_2, user_1);
    // Setup some nodes
    String host_0 = "127.0.0.1";
    FiCaSchedulerNode node_0 = TestUtils.getMockNode(host_0, DEFAULT_RACK, 0, 16 * GB);
    String host_1 = "127.0.0.2";
    FiCaSchedulerNode node_1 = TestUtils.getMockNode(host_1, DEFAULT_RACK, 0, 16 * GB);
    Map<ApplicationAttemptId, FiCaSchedulerApp> apps = ImmutableMap.of(app_0.getApplicationAttemptId(), app_0, app_1.getApplicationAttemptId(), app_1, app_2.getApplicationAttemptId(), app_2);
    Map<NodeId, FiCaSchedulerNode> nodes = ImmutableMap.of(node_0.getNodeID(), node_0, node_1.getNodeID(), node_1);
    final int numNodes = 2;
    Resource clusterResource = Resources.createResource(numNodes * (16 * GB), 1);
    when(csContext.getNumClusterNodes()).thenReturn(numNodes);
    Priority priority = TestUtils.createMockPriority(1);
    app_0.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 1, true, priority, recordFactory)));
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(1 * GB, a.getUsedResources().getMemorySize());
    assertEquals(1 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(0 * GB, app_1.getCurrentConsumption().getMemorySize());
    //Now, headroom is the same for all apps for a given user + queue combo
    //and a change to any app's headroom is reflected for all the user's apps
    //once those apps are active/have themselves calculated headroom for 
    //allocation at least one time
    assertEquals(2 * GB, app_0.getHeadroom().getMemorySize());
    //not yet active
    assertEquals(0 * GB, app_1.getHeadroom().getMemorySize());
    //not yet active
    assertEquals(0 * GB, app_2.getHeadroom().getMemorySize());
    app_1.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority, recordFactory)));
    applyCSAssignment(clusterResource, a.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), a, nodes, apps);
    assertEquals(2 * GB, a.getUsedResources().getMemorySize());
    assertEquals(1 * GB, app_0.getCurrentConsumption().getMemorySize());
    assertEquals(1 * GB, app_1.getCurrentConsumption().getMemorySize());
    assertEquals(1 * GB, app_0.getHeadroom().getMemorySize());
    //now active
    assertEquals(1 * GB, app_1.getHeadroom().getMemorySize());
    //not yet active
    assertEquals(0 * GB, app_2.getHeadroom().getMemorySize());
    //Complete container and verify that headroom is updated, for both apps 
    //for the user
    RMContainer rmContainer = app_0.getLiveContainers().iterator().next();
    a.completedContainer(clusterResource, app_0, node_0, rmContainer, ContainerStatus.newInstance(rmContainer.getContainerId(), ContainerState.COMPLETE, "", ContainerExitStatus.KILLED_BY_RESOURCEMANAGER), RMContainerEventType.KILL, null, true);
    assertEquals(2 * GB, app_0.getHeadroom().getMemorySize());
    assertEquals(2 * GB, app_1.getHeadroom().getMemorySize());
}
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) RMContainer(org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainer) 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) Test(org.junit.Test)

Aggregations

ApplicationAttemptId (org.apache.hadoop.yarn.api.records.ApplicationAttemptId)514 Test (org.junit.Test)362 ApplicationId (org.apache.hadoop.yarn.api.records.ApplicationId)222 ContainerId (org.apache.hadoop.yarn.api.records.ContainerId)170 YarnConfiguration (org.apache.hadoop.yarn.conf.YarnConfiguration)109 RMApp (org.apache.hadoop.yarn.server.resourcemanager.rmapp.RMApp)104 Configuration (org.apache.hadoop.conf.Configuration)87 Resource (org.apache.hadoop.yarn.api.records.Resource)82 ArrayList (java.util.ArrayList)75 NodeId (org.apache.hadoop.yarn.api.records.NodeId)74 NodeAddedSchedulerEvent (org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeAddedSchedulerEvent)65 RMNode (org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode)63 NodeUpdateSchedulerEvent (org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent)60 Path (org.apache.hadoop.fs.Path)55 JobId (org.apache.hadoop.mapreduce.v2.api.records.JobId)53 Priority (org.apache.hadoop.yarn.api.records.Priority)52 RMContainer (org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainer)51 Container (org.apache.hadoop.yarn.api.records.Container)50 FiCaSchedulerApp (org.apache.hadoop.yarn.server.resourcemanager.scheduler.common.fica.FiCaSchedulerApp)49 HashMap (java.util.HashMap)42