use of org.apache.hadoop.yarn.factories.RecordFactory in project hadoop by apache.
the class TestApplicationLifetimeMonitor method testApplicationLifetimeMonitor.
@Test(timeout = 60000)
public void testApplicationLifetimeMonitor() throws Exception {
MockRM rm = null;
try {
rm = new MockRM(conf);
rm.start();
Priority appPriority = Priority.newInstance(0);
MockNM nm1 = rm.registerNode("127.0.0.1:1234", 16 * 1024);
Map<ApplicationTimeoutType, Long> timeouts = new HashMap<ApplicationTimeoutType, Long>();
timeouts.put(ApplicationTimeoutType.LIFETIME, 10L);
RMApp app1 = rm.submitApp(1024, appPriority, timeouts);
// 20L seconds
timeouts.put(ApplicationTimeoutType.LIFETIME, 20L);
RMApp app2 = rm.submitApp(1024, appPriority, timeouts);
nm1.nodeHeartbeat(true);
// Send launch Event
MockAM am1 = rm.sendAMLaunched(app1.getCurrentAppAttempt().getAppAttemptId());
am1.registerAppAttempt();
rm.waitForState(app1.getApplicationId(), RMAppState.KILLED);
Assert.assertTrue("Application killed before lifetime value", (System.currentTimeMillis() - app1.getSubmitTime()) > 10000);
Map<ApplicationTimeoutType, String> updateTimeout = new HashMap<ApplicationTimeoutType, String>();
long newLifetime = 10L;
// update 10L seconds more to timeout
String formatISO8601 = Times.formatISO8601(System.currentTimeMillis() + newLifetime * 1000);
updateTimeout.put(ApplicationTimeoutType.LIFETIME, formatISO8601);
UpdateApplicationTimeoutsRequest request = UpdateApplicationTimeoutsRequest.newInstance(app2.getApplicationId(), updateTimeout);
Map<ApplicationTimeoutType, Long> applicationTimeouts = app2.getApplicationTimeouts();
// has old timeout time
long beforeUpdate = applicationTimeouts.get(ApplicationTimeoutType.LIFETIME);
// update app2 lifetime to new time i.e now + timeout
rm.getRMContext().getClientRMService().updateApplicationTimeouts(request);
applicationTimeouts = app2.getApplicationTimeouts();
long afterUpdate = applicationTimeouts.get(ApplicationTimeoutType.LIFETIME);
Assert.assertTrue("Application lifetime value not updated", afterUpdate > beforeUpdate);
// verify for application report.
RecordFactory recordFactory = RecordFactoryProvider.getRecordFactory(null);
GetApplicationReportRequest appRequest = recordFactory.newRecordInstance(GetApplicationReportRequest.class);
appRequest.setApplicationId(app2.getApplicationId());
Map<ApplicationTimeoutType, ApplicationTimeout> appTimeouts = rm.getRMContext().getClientRMService().getApplicationReport(appRequest).getApplicationReport().getApplicationTimeouts();
Assert.assertTrue("Application Timeout are empty.", !appTimeouts.isEmpty());
ApplicationTimeout timeout = appTimeouts.get(ApplicationTimeoutType.LIFETIME);
Assert.assertEquals("Application timeout string is incorrect.", formatISO8601, timeout.getExpiryTime());
Assert.assertTrue("Application remaining time is incorrect", timeout.getRemainingTime() > 0);
rm.waitForState(app2.getApplicationId(), RMAppState.KILLED);
// verify for app killed with updated lifetime
Assert.assertTrue("Application killed before lifetime value", app2.getFinishTime() > afterUpdate);
} finally {
stopRM(rm);
}
}
use of org.apache.hadoop.yarn.factories.RecordFactory in project hadoop by apache.
the class TestCapacityScheduler method testHeadRoomCalculationWithDRC.
@Test
public void testHeadRoomCalculationWithDRC() throws Exception {
// test with total cluster resource of 20GB memory and 20 vcores.
// the queue where two apps running has user limit 0.8
// allocate 10GB memory and 1 vcore to app 1.
// app 1 should have headroom
// 20GB*0.8 - 10GB = 6GB memory available and 15 vcores.
// allocate 1GB memory and 1 vcore to app2.
// app 2 should have headroom 20GB - 10 - 1 = 1GB memory,
// and 20*0.8 - 1 = 15 vcores.
CapacitySchedulerConfiguration csconf = new CapacitySchedulerConfiguration();
csconf.setResourceComparator(DominantResourceCalculator.class);
YarnConfiguration conf = new YarnConfiguration(csconf);
conf.setClass(YarnConfiguration.RM_SCHEDULER, CapacityScheduler.class, ResourceScheduler.class);
MockRM rm = new MockRM(conf);
rm.start();
CapacityScheduler cs = (CapacityScheduler) rm.getResourceScheduler();
LeafQueue qb = (LeafQueue) cs.getQueue("default");
qb.setUserLimitFactor((float) 0.8);
// add app 1
ApplicationId appId = BuilderUtils.newApplicationId(100, 1);
ApplicationAttemptId appAttemptId = BuilderUtils.newApplicationAttemptId(appId, 1);
RMAppAttemptMetrics attemptMetric = new RMAppAttemptMetrics(appAttemptId, rm.getRMContext());
RMAppImpl app = mock(RMAppImpl.class);
when(app.getApplicationId()).thenReturn(appId);
RMAppAttemptImpl attempt = mock(RMAppAttemptImpl.class);
Container container = mock(Container.class);
when(attempt.getMasterContainer()).thenReturn(container);
ApplicationSubmissionContext submissionContext = mock(ApplicationSubmissionContext.class);
when(attempt.getSubmissionContext()).thenReturn(submissionContext);
when(attempt.getAppAttemptId()).thenReturn(appAttemptId);
when(attempt.getRMAppAttemptMetrics()).thenReturn(attemptMetric);
when(app.getCurrentAppAttempt()).thenReturn(attempt);
rm.getRMContext().getRMApps().put(appId, app);
SchedulerEvent addAppEvent = new AppAddedSchedulerEvent(appId, "default", "user1");
cs.handle(addAppEvent);
SchedulerEvent addAttemptEvent = new AppAttemptAddedSchedulerEvent(appAttemptId, false);
cs.handle(addAttemptEvent);
// add app 2
ApplicationId appId2 = BuilderUtils.newApplicationId(100, 2);
ApplicationAttemptId appAttemptId2 = BuilderUtils.newApplicationAttemptId(appId2, 1);
RMAppAttemptMetrics attemptMetric2 = new RMAppAttemptMetrics(appAttemptId2, rm.getRMContext());
RMAppImpl app2 = mock(RMAppImpl.class);
when(app2.getApplicationId()).thenReturn(appId2);
RMAppAttemptImpl attempt2 = mock(RMAppAttemptImpl.class);
when(attempt2.getMasterContainer()).thenReturn(container);
when(attempt2.getSubmissionContext()).thenReturn(submissionContext);
when(attempt2.getAppAttemptId()).thenReturn(appAttemptId2);
when(attempt2.getRMAppAttemptMetrics()).thenReturn(attemptMetric2);
when(app2.getCurrentAppAttempt()).thenReturn(attempt2);
rm.getRMContext().getRMApps().put(appId2, app2);
addAppEvent = new AppAddedSchedulerEvent(appId2, "default", "user2");
cs.handle(addAppEvent);
addAttemptEvent = new AppAttemptAddedSchedulerEvent(appAttemptId2, false);
cs.handle(addAttemptEvent);
// add nodes to cluster, so cluster have 20GB and 20 vcores
Resource newResource = Resource.newInstance(10 * GB, 10);
RMNode node = MockNodes.newNodeInfo(0, newResource, 1, "127.0.0.1");
cs.handle(new NodeAddedSchedulerEvent(node));
Resource newResource2 = Resource.newInstance(10 * GB, 10);
RMNode node2 = MockNodes.newNodeInfo(0, newResource2, 1, "127.0.0.2");
cs.handle(new NodeAddedSchedulerEvent(node2));
FiCaSchedulerApp fiCaApp1 = cs.getSchedulerApplications().get(app.getApplicationId()).getCurrentAppAttempt();
FiCaSchedulerApp fiCaApp2 = cs.getSchedulerApplications().get(app2.getApplicationId()).getCurrentAppAttempt();
Priority u0Priority = TestUtils.createMockPriority(1);
RecordFactory recordFactory = RecordFactoryProvider.getRecordFactory(null);
// allocate container for app1 with 10GB memory and 1 vcore
fiCaApp1.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 10 * GB, 1, true, u0Priority, recordFactory)));
cs.handle(new NodeUpdateSchedulerEvent(node));
cs.handle(new NodeUpdateSchedulerEvent(node2));
assertEquals(6 * GB, fiCaApp1.getHeadroom().getMemorySize());
assertEquals(15, fiCaApp1.getHeadroom().getVirtualCores());
// allocate container for app2 with 1GB memory and 1 vcore
fiCaApp2.updateResourceRequests(Collections.singletonList(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 1, true, u0Priority, recordFactory)));
cs.handle(new NodeUpdateSchedulerEvent(node));
cs.handle(new NodeUpdateSchedulerEvent(node2));
assertEquals(9 * GB, fiCaApp2.getHeadroom().getMemorySize());
assertEquals(15, fiCaApp2.getHeadroom().getVirtualCores());
}
use of org.apache.hadoop.yarn.factories.RecordFactory in project hadoop by apache.
the class TestApplicationLimitsByPartition method testHeadroom.
@Test
public void testHeadroom() throws Exception {
/*
* Test Case: Verify Headroom calculated is sum of headrooms for each
* partition requested. So submit a app with requests for default partition
* and 'x' partition, so the total headroom for the user should be sum of
* the head room for both labels.
*/
simpleNodeLabelMappingToManager();
CapacitySchedulerConfiguration csConf = (CapacitySchedulerConfiguration) TestUtils.getComplexConfigurationWithQueueLabels(conf);
final String A1 = CapacitySchedulerConfiguration.ROOT + ".a" + ".a1";
final String B2 = CapacitySchedulerConfiguration.ROOT + ".b" + ".b2";
csConf.setUserLimit(A1, 25);
csConf.setUserLimit(B2, 25);
YarnConfiguration conf = new YarnConfiguration();
CapacitySchedulerContext csContext = mock(CapacitySchedulerContext.class);
when(csContext.getConfiguration()).thenReturn(csConf);
when(csContext.getConf()).thenReturn(conf);
when(csContext.getMinimumResourceCapability()).thenReturn(Resources.createResource(GB));
when(csContext.getMaximumResourceCapability()).thenReturn(Resources.createResource(16 * GB));
when(csContext.getResourceCalculator()).thenReturn(resourceCalculator);
RMContext rmContext = TestUtils.getMockRMContext();
RMContext spyRMContext = spy(rmContext);
when(spyRMContext.getNodeLabelManager()).thenReturn(mgr);
when(csContext.getRMContext()).thenReturn(spyRMContext);
mgr.activateNode(NodeId.newInstance("h0", 0), // default Label
Resource.newInstance(160 * GB, 16));
mgr.activateNode(NodeId.newInstance("h1", 0), // label x
Resource.newInstance(160 * GB, 16));
mgr.activateNode(NodeId.newInstance("h2", 0), // label y
Resource.newInstance(160 * GB, 16));
// Say cluster has 100 nodes of 16G each
Resource clusterResource = Resources.createResource(160 * GB);
when(csContext.getClusterResource()).thenReturn(clusterResource);
Map<String, CSQueue> queues = new HashMap<String, CSQueue>();
CSQueue rootQueue = CapacitySchedulerQueueManager.parseQueue(csContext, csConf, null, "root", queues, queues, TestUtils.spyHook);
ResourceUsage queueResUsage = rootQueue.getQueueResourceUsage();
when(csContext.getClusterResourceUsage()).thenReturn(queueResUsage);
// Manipulate queue 'a'
LeafQueue queue = TestLeafQueue.stubLeafQueue((LeafQueue) queues.get("b2"));
queue.updateClusterResource(clusterResource, new ResourceLimits(clusterResource));
String rack_0 = "rack_0";
FiCaSchedulerNode node_0 = TestUtils.getMockNode("h0", rack_0, 0, 160 * GB);
FiCaSchedulerNode node_1 = TestUtils.getMockNode("h1", rack_0, 0, 160 * GB);
final String user_0 = "user_0";
final String user_1 = "user_1";
RecordFactory recordFactory = RecordFactoryProvider.getRecordFactory(null);
ConcurrentMap<ApplicationId, RMApp> spyApps = spy(new ConcurrentHashMap<ApplicationId, RMApp>());
RMApp rmApp = mock(RMApp.class);
ResourceRequest amResourceRequest = mock(ResourceRequest.class);
Resource amResource = Resources.createResource(0, 0);
when(amResourceRequest.getCapability()).thenReturn(amResource);
when(rmApp.getAMResourceRequest()).thenReturn(amResourceRequest);
Mockito.doReturn(rmApp).when(spyApps).get((ApplicationId) Matchers.any());
when(spyRMContext.getRMApps()).thenReturn(spyApps);
RMAppAttempt rmAppAttempt = mock(RMAppAttempt.class);
when(rmApp.getRMAppAttempt((ApplicationAttemptId) Matchers.any())).thenReturn(rmAppAttempt);
when(rmApp.getCurrentAppAttempt()).thenReturn(rmAppAttempt);
Mockito.doReturn(rmApp).when(spyApps).get((ApplicationId) Matchers.any());
Mockito.doReturn(true).when(spyApps).containsKey((ApplicationId) Matchers.any());
Priority priority_1 = TestUtils.createMockPriority(1);
// Submit first application with some resource-requests from user_0,
// and check headroom
final ApplicationAttemptId appAttemptId_0_0 = TestUtils.getMockApplicationAttemptId(0, 0);
FiCaSchedulerApp app_0_0 = new FiCaSchedulerApp(appAttemptId_0_0, user_0, queue, queue.getAbstractUsersManager(), spyRMContext);
queue.submitApplicationAttempt(app_0_0, user_0);
List<ResourceRequest> app_0_0_requests = new ArrayList<ResourceRequest>();
app_0_0_requests.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority_1, recordFactory));
app_0_0.updateResourceRequests(app_0_0_requests);
// Schedule to compute
queue.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
//head room = queue capacity = 50 % 90% 160 GB * 0.25 (UL)
Resource expectedHeadroom = Resources.createResource((int) (0.5 * 0.9 * 160 * 0.25) * GB, 1);
assertEquals(expectedHeadroom, app_0_0.getHeadroom());
// Submit second application from user_0, check headroom
final ApplicationAttemptId appAttemptId_0_1 = TestUtils.getMockApplicationAttemptId(1, 0);
FiCaSchedulerApp app_0_1 = new FiCaSchedulerApp(appAttemptId_0_1, user_0, queue, queue.getAbstractUsersManager(), spyRMContext);
queue.submitApplicationAttempt(app_0_1, user_0);
List<ResourceRequest> app_0_1_requests = new ArrayList<ResourceRequest>();
app_0_1_requests.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority_1, recordFactory));
app_0_1.updateResourceRequests(app_0_1_requests);
app_0_1_requests.clear();
app_0_1_requests.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority_1, recordFactory, "y"));
app_0_1.updateResourceRequests(app_0_1_requests);
// Schedule to compute
queue.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), // Schedule to compute
SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
queue.assignContainers(clusterResource, node_1, new ResourceLimits(clusterResource), // Schedule to compute
SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
// no change
assertEquals(expectedHeadroom, app_0_0.getHeadroom());
//head room for default label + head room for y partition
//head room for y partition = 100% 50%(b queue capacity ) * 160 * GB
Resource expectedHeadroomWithReqInY = Resources.add(Resources.createResource((int) (0.25 * 0.5 * 160) * GB, 1), expectedHeadroom);
assertEquals(expectedHeadroomWithReqInY, app_0_1.getHeadroom());
// Submit first application from user_1, check for new headroom
final ApplicationAttemptId appAttemptId_1_0 = TestUtils.getMockApplicationAttemptId(2, 0);
FiCaSchedulerApp app_1_0 = new FiCaSchedulerApp(appAttemptId_1_0, user_1, queue, queue.getAbstractUsersManager(), spyRMContext);
queue.submitApplicationAttempt(app_1_0, user_1);
List<ResourceRequest> app_1_0_requests = new ArrayList<ResourceRequest>();
app_1_0_requests.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority_1, recordFactory));
app_1_0.updateResourceRequests(app_1_0_requests);
app_1_0_requests.clear();
app_1_0_requests.add(TestUtils.createResourceRequest(ResourceRequest.ANY, 1 * GB, 2, true, priority_1, recordFactory, "y"));
app_1_0.updateResourceRequests(app_1_0_requests);
// Schedule to compute
queue.assignContainers(clusterResource, node_0, new ResourceLimits(clusterResource), // Schedule to compute
SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
//head room = queue capacity = (50 % 90% 160 GB)/2 (for 2 users)
expectedHeadroom = Resources.createResource((int) (0.5 * 0.9 * 160 * 0.25) * GB, 1);
//head room for default label + head room for y partition
//head room for y partition = 100% 50%(b queue capacity ) * 160 * GB
expectedHeadroomWithReqInY = Resources.add(Resources.createResource((int) (0.25 * 0.5 * 160) * GB, 1), expectedHeadroom);
assertEquals(expectedHeadroom, app_0_0.getHeadroom());
assertEquals(expectedHeadroomWithReqInY, app_0_1.getHeadroom());
assertEquals(expectedHeadroomWithReqInY, app_1_0.getHeadroom());
}
use of org.apache.hadoop.yarn.factories.RecordFactory in project hadoop by apache.
the class NodeManager method createNodeStatus.
public static org.apache.hadoop.yarn.server.api.records.NodeStatus createNodeStatus(NodeId nodeId, List<ContainerStatus> containers) {
RecordFactory recordFactory = RecordFactoryProvider.getRecordFactory(null);
org.apache.hadoop.yarn.server.api.records.NodeStatus nodeStatus = recordFactory.newRecordInstance(org.apache.hadoop.yarn.server.api.records.NodeStatus.class);
nodeStatus.setNodeId(nodeId);
nodeStatus.setContainersStatuses(containers);
NodeHealthStatus nodeHealthStatus = recordFactory.newRecordInstance(NodeHealthStatus.class);
nodeHealthStatus.setIsNodeHealthy(true);
nodeStatus.setNodeHealthStatus(nodeHealthStatus);
return nodeStatus;
}
use of org.apache.hadoop.yarn.factories.RecordFactory in project hadoop by apache.
the class TestTaskAttemptListenerImpl method createTce.
private static TaskAttemptCompletionEvent createTce(int eventId, boolean isMap, TaskAttemptCompletionEventStatus status) {
JobId jid = MRBuilderUtils.newJobId(12345, 1, 1);
TaskId tid = MRBuilderUtils.newTaskId(jid, 0, isMap ? org.apache.hadoop.mapreduce.v2.api.records.TaskType.MAP : org.apache.hadoop.mapreduce.v2.api.records.TaskType.REDUCE);
TaskAttemptId attemptId = MRBuilderUtils.newTaskAttemptId(tid, 0);
RecordFactory recordFactory = RecordFactoryProvider.getRecordFactory(null);
TaskAttemptCompletionEvent tce = recordFactory.newRecordInstance(TaskAttemptCompletionEvent.class);
tce.setEventId(eventId);
tce.setAttemptId(attemptId);
tce.setStatus(status);
return tce;
}
Aggregations