YARN-4344. NMs reconnecting with changed capabilities can lead to wrong cluster resource calculations. Contributed by Varun Vasudev

(cherry picked from commit d36b6e045f)

Conflicts:

	hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/TestCapacityScheduler.java
This commit is contained in:
Jason Lowe 2015-11-23 20:47:12 +00:00
parent f62cdc38f3
commit 5bd40d7099
4 changed files with 123 additions and 6 deletions

View File

@ -28,6 +28,9 @@ Release 2.7.3 - UNRELEASED
YARN-3769. Consider user limit when calculating total pending resource for
preemption policy in Capacity Scheduler. (Eric Payne via wangda)
YARN-4344. NMs reconnecting with changed capabilities can lead to wrong
cluster resource calculations (Varun Vasudev via jlowe)
Release 2.7.2 - UNRELEASED
INCOMPATIBLE CHANGES
@ -908,6 +911,9 @@ Release 2.6.3 - UNRELEASED
YARN-2859. ApplicationHistoryServer binds to default port 8188 in MiniYARNCluster.
(Vinod Kumar Vavilapalli via xgong)
YARN-4344. NMs reconnecting with changed capabilities can lead to wrong
cluster resource calculations (Varun Vasudev via jlowe)
Release 2.6.2 - 2015-10-28
INCOMPATIBLE CHANGES

View File

@ -1319,12 +1319,12 @@ public class CapacityScheduler extends
FiCaSchedulerNode schedulerNode = new FiCaSchedulerNode(nodeManager,
usePortForNodeName, nodeManager.getNodeLabels());
this.nodes.put(nodeManager.getNodeID(), schedulerNode);
Resources.addTo(clusterResource, nodeManager.getTotalCapability());
Resources.addTo(clusterResource, schedulerNode.getTotalResource());
// update this node to node label manager
if (labelManager != null) {
labelManager.activateNode(nodeManager.getNodeID(),
nodeManager.getTotalCapability());
schedulerNode.getTotalResource());
}
root.updateClusterResource(clusterResource, new ResourceLimits(
@ -1350,7 +1350,7 @@ public class CapacityScheduler extends
if (node == null) {
return;
}
Resources.subtractFrom(clusterResource, node.getRMNode().getTotalCapability());
Resources.subtractFrom(clusterResource, node.getTotalResource());
root.updateClusterResource(clusterResource, new ResourceLimits(
clusterResource));
int numNodes = numNodeManagers.decrementAndGet();

View File

@ -915,7 +915,7 @@ public class FifoScheduler extends
updateMaximumAllocation(node, false);
// Update cluster metrics
Resources.subtractFrom(clusterResource, node.getRMNode().getTotalCapability());
Resources.subtractFrom(clusterResource, node.getTotalResource());
}
@Override
@ -938,7 +938,7 @@ public class FifoScheduler extends
FiCaSchedulerNode schedulerNode = new FiCaSchedulerNode(nodeManager,
usePortForNodeName);
this.nodes.put(nodeManager.getNodeID(), schedulerNode);
Resources.addTo(clusterResource, nodeManager.getTotalCapability());
Resources.addTo(clusterResource, schedulerNode.getTotalResource());
updateMaximumAllocation(schedulerNode, true);
}

View File

@ -45,6 +45,7 @@ import org.apache.hadoop.security.Credentials;
import org.apache.hadoop.security.UserGroupInformation;
import org.apache.hadoop.security.token.Token;
import org.apache.hadoop.security.token.TokenIdentifier;
import org.apache.hadoop.service.Service;
import org.apache.hadoop.yarn.LocalConfigurationProvider;
import org.apache.hadoop.yarn.api.ApplicationMasterProtocol;
import org.apache.hadoop.yarn.api.protocolrecords.AllocateRequest;
@ -68,11 +69,15 @@ import org.apache.hadoop.yarn.api.records.ResourceOption;
import org.apache.hadoop.yarn.api.records.ResourceRequest;
import org.apache.hadoop.yarn.conf.YarnConfiguration;
import org.apache.hadoop.yarn.event.AsyncDispatcher;
import org.apache.hadoop.yarn.event.Dispatcher;
import org.apache.hadoop.yarn.event.DrainDispatcher;
import org.apache.hadoop.yarn.event.EventHandler;
import org.apache.hadoop.yarn.exceptions.YarnException;
import org.apache.hadoop.yarn.exceptions.YarnRuntimeException;
import org.apache.hadoop.yarn.factories.RecordFactory;
import org.apache.hadoop.yarn.factory.providers.RecordFactoryProvider;
import org.apache.hadoop.yarn.ipc.YarnRPC;
import org.apache.hadoop.yarn.server.api.protocolrecords.RegisterNodeManagerRequest;
import org.apache.hadoop.yarn.server.api.protocolrecords.UpdateNodeResourceRequest;
import org.apache.hadoop.yarn.server.resourcemanager.AdminService;
import org.apache.hadoop.yarn.server.resourcemanager.Application;
@ -80,10 +85,13 @@ import org.apache.hadoop.yarn.server.resourcemanager.MockAM;
import org.apache.hadoop.yarn.server.resourcemanager.MockNM;
import org.apache.hadoop.yarn.server.resourcemanager.MockNodes;
import org.apache.hadoop.yarn.server.resourcemanager.MockRM;
import org.apache.hadoop.yarn.server.resourcemanager.NMLivelinessMonitor;
import org.apache.hadoop.yarn.server.resourcemanager.NodeManager;
import org.apache.hadoop.yarn.server.resourcemanager.NodesListManager;
import org.apache.hadoop.yarn.server.resourcemanager.RMContext;
import org.apache.hadoop.yarn.server.resourcemanager.RMContextImpl;
import org.apache.hadoop.yarn.server.resourcemanager.ResourceManager;
import org.apache.hadoop.yarn.server.resourcemanager.ResourceTrackerService;
import org.apache.hadoop.yarn.server.resourcemanager.Task;
import org.apache.hadoop.yarn.server.resourcemanager.TestAMAuthorization.MockRMWithAMS;
import org.apache.hadoop.yarn.server.resourcemanager.TestAMAuthorization.MyContainerManager;
@ -101,6 +109,7 @@ import org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptS
import org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainer;
import org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainerState;
import org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode;
import org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNodeEventType;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.AbstractYarnScheduler;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.Allocation;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.ResourceScheduler;
@ -117,6 +126,7 @@ import org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeAddedSc
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeRemovedSchedulerEvent;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.NodeUpdateSchedulerEvent;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.SchedulerEvent;
import org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.SchedulerEventType;
import org.apache.hadoop.yarn.server.resourcemanager.security.ClientToAMTokenSecretManagerInRM;
import org.apache.hadoop.yarn.server.resourcemanager.security.NMTokenSecretManagerInRM;
import org.apache.hadoop.yarn.server.resourcemanager.security.RMContainerTokenSecretManager;
@ -135,7 +145,6 @@ import org.junit.ComparisonFailure;
import org.junit.Test;
import org.mockito.Mockito;
public class TestCapacityScheduler {
private static final Log LOG = LogFactory.getLog(TestCapacityScheduler.class);
private final int GB = 1024;
@ -2831,4 +2840,106 @@ public class TestCapacityScheduler {
+ CapacitySchedulerConfiguration.MAXIMUM_ALLOCATION_VCORES;
conf.setInt(propName, maxAllocVcores);
}
private class SleepHandler implements EventHandler<SchedulerEvent> {
boolean sleepFlag = false;
int sleepTime = 20;
@Override
public void handle(SchedulerEvent event) {
try {
if(sleepFlag) {
Thread.sleep(sleepTime);
}
}
catch(InterruptedException ie) {
}
}
}
private ResourceTrackerService getPrivateResourceTrackerService(
Dispatcher privateDispatcher, SleepHandler sleepHandler) {
Configuration conf = new Configuration();
ResourceTrackerService privateResourceTrackerService;
RMContext privateContext =
new RMContextImpl(privateDispatcher, null, null, null, null, null, null,
null, null, null);
privateContext.setNodeLabelManager(Mockito.mock(RMNodeLabelsManager.class));
privateDispatcher.register(SchedulerEventType.class, sleepHandler);
privateDispatcher.register(SchedulerEventType.class,
resourceManager.getResourceScheduler());
privateDispatcher.register(RMNodeEventType.class,
new ResourceManager.NodeEventDispatcher(privateContext));
((Service) privateDispatcher).init(conf);
((Service) privateDispatcher).start();
NMLivelinessMonitor nmLivelinessMonitor =
new NMLivelinessMonitor(privateDispatcher);
nmLivelinessMonitor.init(conf);
nmLivelinessMonitor.start();
NodesListManager nodesListManager = new NodesListManager(privateContext);
nodesListManager.init(conf);
RMContainerTokenSecretManager containerTokenSecretManager =
new RMContainerTokenSecretManager(conf);
containerTokenSecretManager.start();
NMTokenSecretManagerInRM nmTokenSecretManager =
new NMTokenSecretManagerInRM(conf);
nmTokenSecretManager.start();
privateResourceTrackerService =
new ResourceTrackerService(privateContext, nodesListManager,
nmLivelinessMonitor, containerTokenSecretManager,
nmTokenSecretManager);
privateResourceTrackerService.init(conf);
privateResourceTrackerService.start();
resourceManager.getResourceScheduler().setRMContext(privateContext);
return privateResourceTrackerService;
}
/**
* Test the behaviour of the capacity scheduler when a node reconnects
* with changed capabilities. This test is to catch any race conditions
* that might occur due to the use of the RMNode object.
* @throws Exception
*/
@Test
public void testNodemanagerReconnect() throws Exception {
DrainDispatcher privateDispatcher = new DrainDispatcher();
SleepHandler sleepHandler = new SleepHandler();
ResourceTrackerService privateResourceTrackerService =
getPrivateResourceTrackerService(privateDispatcher,
sleepHandler);
// Register node1
String hostname1 = "localhost1";
Resource capability = BuilderUtils.newResource(4096, 4);
RecordFactory recordFactory = RecordFactoryProvider.getRecordFactory(null);
RegisterNodeManagerRequest request1 =
recordFactory.newRecordInstance(RegisterNodeManagerRequest.class);
NodeId nodeId1 = NodeId.newInstance(hostname1, 0);
request1.setNodeId(nodeId1);
request1.setHttpPort(0);
request1.setResource(capability);
privateResourceTrackerService.registerNodeManager(request1);
privateDispatcher.await();
Resource clusterResource = resourceManager.getResourceScheduler().getClusterResource();
Assert.assertEquals("Initial cluster resources don't match", capability,
clusterResource);
Resource newCapability = BuilderUtils.newResource(1024, 1);
RegisterNodeManagerRequest request2 =
recordFactory.newRecordInstance(RegisterNodeManagerRequest.class);
request2.setNodeId(nodeId1);
request2.setHttpPort(0);
request2.setResource(newCapability);
// hold up the disaptcher and register the same node with lower capability
sleepHandler.sleepFlag = true;
privateResourceTrackerService.registerNodeManager(request2);
privateDispatcher.await();
Assert.assertEquals("Cluster resources don't match", newCapability,
resourceManager.getResourceScheduler().getClusterResource());
privateResourceTrackerService.stop();
}
}