use of org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest in project hadoop by apache.
the class TestAMRMClientOnRMRestart method testAMRMClientForUnregisterAMOnRMRestart.
// Test verify for
// 1. AM try to unregister without registering
// 2. AM register to RM, and try to unregister immediately after RM restart
@Test(timeout = 60000)
public void testAMRMClientForUnregisterAMOnRMRestart() throws Exception {
MemoryRMStateStore memStore = new MemoryRMStateStore();
memStore.init(conf);
// Phase-1 Start 1st RM
MyResourceManager rm1 = new MyResourceManager(conf, memStore);
rm1.start();
DrainDispatcher dispatcher = (DrainDispatcher) rm1.getRMContext().getDispatcher();
// Submit the application
RMApp app = rm1.submitApp(1024);
dispatcher.await();
MockNM nm1 = new MockNM("h1:1234", 15120, rm1.getResourceTrackerService());
nm1.registerNode();
// Node heartbeat
nm1.nodeHeartbeat(true);
dispatcher.await();
ApplicationAttemptId appAttemptId = app.getCurrentAppAttempt().getAppAttemptId();
rm1.sendAMLaunched(appAttemptId);
dispatcher.await();
org.apache.hadoop.security.token.Token<AMRMTokenIdentifier> token = rm1.getRMContext().getRMApps().get(appAttemptId.getApplicationId()).getRMAppAttempt(appAttemptId).getAMRMToken();
UserGroupInformation ugi = UserGroupInformation.getCurrentUser();
ugi.addTokenIdentifier(token.decodeIdentifier());
AMRMClient<ContainerRequest> amClient = new MyAMRMClientImpl(rm1);
amClient.init(conf);
amClient.start();
amClient.registerApplicationMaster("h1", 10000, "");
amClient.allocate(0.1f);
// Phase-2 start 2nd RM is up
MyResourceManager rm2 = new MyResourceManager(conf, memStore);
rm2.start();
nm1.setResourceTrackerService(rm2.getResourceTrackerService());
((MyAMRMClientImpl) amClient).updateRMProxy(rm2);
dispatcher = (DrainDispatcher) rm2.getRMContext().getDispatcher();
// NM should be rebooted on heartbeat, even first heartbeat for nm2
NodeHeartbeatResponse hbResponse = nm1.nodeHeartbeat(true);
Assert.assertEquals(NodeAction.RESYNC, hbResponse.getNodeAction());
// new NM to represent NM re-register
nm1 = new MockNM("h1:1234", 10240, rm2.getResourceTrackerService());
ContainerId containerId = ContainerId.newContainerId(appAttemptId, 1);
NMContainerStatus containerReport = NMContainerStatus.newInstance(containerId, 0, ContainerState.RUNNING, Resource.newInstance(1024, 1), "recover container", 0, Priority.newInstance(0), 0);
nm1.registerNode(Arrays.asList(containerReport), null);
nm1.nodeHeartbeat(true);
dispatcher.await();
amClient.unregisterApplicationMaster(FinalApplicationStatus.SUCCEEDED, null, null);
rm2.waitForState(appAttemptId, RMAppAttemptState.FINISHING);
nm1.nodeHeartbeat(appAttemptId, 1, ContainerState.COMPLETE);
rm2.waitForState(appAttemptId, RMAppAttemptState.FINISHED);
rm2.waitForState(app.getApplicationId(), RMAppState.FINISHED);
amClient.stop();
rm1.stop();
rm2.stop();
}
use of org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest in project hadoop by apache.
the class TestDistributedScheduling method testAMOpportunistic.
/**
* Check if an AM can ask for opportunistic containers and get them.
* @throws Exception
*/
@Test
public void testAMOpportunistic() throws Exception {
// Basic container to request
Resource capability = Resource.newInstance(1024, 1);
Priority priority = Priority.newInstance(1);
// Get the cluster topology
List<NodeReport> nodeReports = rmClient.getNodeReports(NodeState.RUNNING);
String node = nodeReports.get(0).getNodeId().getHost();
String rack = nodeReports.get(0).getRackName();
String[] nodes = new String[] { node };
String[] racks = new String[] { rack };
// Create an AM to request resources
AMRMClient<AMRMClient.ContainerRequest> amClient = null;
try {
amClient = new AMRMClientImpl<AMRMClient.ContainerRequest>(client);
amClient.init(yarnConf);
amClient.start();
amClient.registerApplicationMaster(NetUtils.getHostname(), 1024, "");
// AM requests an opportunistic container
ExecutionTypeRequest execTypeRequest = ExecutionTypeRequest.newInstance(ExecutionType.OPPORTUNISTIC, true);
ContainerRequest containerRequest = new AMRMClient.ContainerRequest(capability, nodes, racks, priority, 0, true, null, execTypeRequest);
amClient.addContainerRequest(containerRequest);
// Wait until the container is allocated
ContainerId opportunisticContainerId = null;
for (int i = 0; i < 10 && opportunisticContainerId == null; i++) {
AllocateResponse allocResponse = amClient.allocate(0.1f);
List<Container> allocatedContainers = allocResponse.getAllocatedContainers();
for (Container allocatedContainer : allocatedContainers) {
// Check that this is the container we required
assertEquals(ExecutionType.OPPORTUNISTIC, allocatedContainer.getExecutionType());
opportunisticContainerId = allocatedContainer.getId();
}
sleep(100);
}
assertNotNull(opportunisticContainerId);
// The RM sees the container as OPPORTUNISTIC
ResourceScheduler scheduler = cluster.getResourceManager().getResourceScheduler();
RMContainer rmContainer = scheduler.getRMContainer(opportunisticContainerId);
assertEquals(ExecutionType.OPPORTUNISTIC, rmContainer.getExecutionType());
// Release the opportunistic container
amClient.releaseAssignedContainer(opportunisticContainerId);
// Wait for the release container to appear
boolean released = false;
for (int i = 0; i < 10 && !released; i++) {
AllocateResponse allocResponse = amClient.allocate(0.1f);
List<ContainerStatus> completedContainers = allocResponse.getCompletedContainersStatuses();
for (ContainerStatus completedContainer : completedContainers) {
ContainerId completedContainerId = completedContainer.getContainerId();
assertEquals(completedContainerId, opportunisticContainerId);
released = true;
}
if (!released) {
sleep(100);
}
}
assertTrue(released);
// The RM shouldn't see the container anymore
rmContainer = scheduler.getRMContainer(opportunisticContainerId);
assertNull(rmContainer);
// Clean the AM
amClient.unregisterApplicationMaster(FinalApplicationStatus.SUCCEEDED, null, null);
} finally {
if (amClient != null && amClient.getServiceState() == Service.STATE.STARTED) {
amClient.close();
}
}
}
use of org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest in project hadoop by apache.
the class TestDistributedScheduling method testAMRMClient.
/**
* Validates if AMRMClient can be used with Distributed Scheduling turned on.
*
* @throws Exception
*/
@Test(timeout = 120000)
@SuppressWarnings("unchecked")
public void testAMRMClient() throws Exception {
AMRMClientImpl<AMRMClient.ContainerRequest> amClient = null;
try {
Priority priority = Priority.newInstance(1);
Priority priority2 = Priority.newInstance(2);
Resource capability = Resource.newInstance(1024, 1);
List<NodeReport> nodeReports = rmClient.getNodeReports(NodeState.RUNNING);
String node = nodeReports.get(0).getNodeId().getHost();
String rack = nodeReports.get(0).getRackName();
String[] nodes = new String[] { node };
String[] racks = new String[] { rack };
// start am rm client
amClient = new AMRMClientImpl(client);
amClient.init(yarnConf);
amClient.start();
amClient.registerApplicationMaster(NetUtils.getHostname(), 1024, "");
assertEquals(0, amClient.ask.size());
assertEquals(0, amClient.release.size());
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, null, null, priority2, 0, true, null, ExecutionTypeRequest.newInstance(ExecutionType.OPPORTUNISTIC, true)));
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, null, null, priority2, 0, true, null, ExecutionTypeRequest.newInstance(ExecutionType.OPPORTUNISTIC, true)));
amClient.removeContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.removeContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.removeContainerRequest(new AMRMClient.ContainerRequest(capability, null, null, priority2, 0, true, null, ExecutionTypeRequest.newInstance(ExecutionType.OPPORTUNISTIC, true)));
RemoteRequestsTable<ContainerRequest> remoteRequestsTable = amClient.getTable(0);
int containersRequestedNode = remoteRequestsTable.get(priority, node, ExecutionType.GUARANTEED, capability).remoteRequest.getNumContainers();
int containersRequestedRack = remoteRequestsTable.get(priority, rack, ExecutionType.GUARANTEED, capability).remoteRequest.getNumContainers();
int containersRequestedAny = remoteRequestsTable.get(priority, ResourceRequest.ANY, ExecutionType.GUARANTEED, capability).remoteRequest.getNumContainers();
int oppContainersRequestedAny = remoteRequestsTable.get(priority2, ResourceRequest.ANY, ExecutionType.OPPORTUNISTIC, capability).remoteRequest.getNumContainers();
assertEquals(2, containersRequestedNode);
assertEquals(2, containersRequestedRack);
assertEquals(2, containersRequestedAny);
assertEquals(1, oppContainersRequestedAny);
assertEquals(4, amClient.ask.size());
assertEquals(0, amClient.release.size());
// RM should allocate container within 2 calls to allocate()
int allocatedContainerCount = 0;
int iterationsLeft = 10;
Set<ContainerId> releases = new TreeSet<>();
amClient.getNMTokenCache().clearCache();
Assert.assertEquals(0, amClient.getNMTokenCache().numberOfTokensInCache());
HashMap<String, Token> receivedNMTokens = new HashMap<>();
while (allocatedContainerCount < (containersRequestedAny + oppContainersRequestedAny) && iterationsLeft-- > 0) {
AllocateResponse allocResponse = amClient.allocate(0.1f);
assertEquals(0, amClient.ask.size());
assertEquals(0, amClient.release.size());
allocatedContainerCount += allocResponse.getAllocatedContainers().size();
for (Container container : allocResponse.getAllocatedContainers()) {
ContainerId rejectContainerId = container.getId();
releases.add(rejectContainerId);
}
for (NMToken token : allocResponse.getNMTokens()) {
String nodeID = token.getNodeId().toString();
receivedNMTokens.put(nodeID, token.getToken());
}
if (allocatedContainerCount < containersRequestedAny) {
// sleep to let NM's heartbeat to RM and trigger allocations
sleep(100);
}
}
assertEquals(allocatedContainerCount, containersRequestedAny + oppContainersRequestedAny);
for (ContainerId rejectContainerId : releases) {
amClient.releaseAssignedContainer(rejectContainerId);
}
assertEquals(3, amClient.release.size());
assertEquals(0, amClient.ask.size());
// need to tell the AMRMClient that we dont need these resources anymore
amClient.removeContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.removeContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.removeContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority2, 0, true, null, ExecutionTypeRequest.newInstance(ExecutionType.OPPORTUNISTIC, true)));
assertEquals(4, amClient.ask.size());
// test RPC exception handling
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amClient.addContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority2, 0, true, null, ExecutionTypeRequest.newInstance(ExecutionType.OPPORTUNISTIC, true)));
final AMRMClient amc = amClient;
ApplicationMasterProtocol realRM = amClient.rmClient;
try {
ApplicationMasterProtocol mockRM = mock(ApplicationMasterProtocol.class);
when(mockRM.allocate(any(AllocateRequest.class))).thenAnswer(new Answer<AllocateResponse>() {
public AllocateResponse answer(InvocationOnMock invocation) throws Exception {
amc.removeContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amc.removeContainerRequest(new AMRMClient.ContainerRequest(capability, nodes, racks, priority));
amc.removeContainerRequest(new AMRMClient.ContainerRequest(capability, null, null, priority2, 0, true, null, ExecutionTypeRequest.newInstance(ExecutionType.OPPORTUNISTIC, true)));
throw new Exception();
}
});
amClient.rmClient = mockRM;
amClient.allocate(0.1f);
} catch (Exception ioe) {
} finally {
amClient.rmClient = realRM;
}
assertEquals(3, amClient.release.size());
assertEquals(6, amClient.ask.size());
iterationsLeft = 3;
// do a few iterations to ensure RM is not going send new containers
while (iterationsLeft-- > 0) {
// inform RM of rejection
AllocateResponse allocResponse = amClient.allocate(0.1f);
// RM did not send new containers because AM does not need any
assertEquals(0, allocResponse.getAllocatedContainers().size());
if (allocResponse.getCompletedContainersStatuses().size() > 0) {
for (ContainerStatus cStatus : allocResponse.getCompletedContainersStatuses()) {
if (releases.contains(cStatus.getContainerId())) {
assertEquals(cStatus.getState(), ContainerState.COMPLETE);
assertEquals(-100, cStatus.getExitStatus());
releases.remove(cStatus.getContainerId());
}
}
}
if (iterationsLeft > 0) {
// sleep to make sure NM's heartbeat
sleep(100);
}
}
assertEquals(0, amClient.ask.size());
assertEquals(0, amClient.release.size());
amClient.unregisterApplicationMaster(FinalApplicationStatus.SUCCEEDED, null, null);
} finally {
if (amClient != null && amClient.getServiceState() == Service.STATE.STARTED) {
amClient.stop();
}
}
}
use of org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest in project hadoop by apache.
the class ApplicationMaster method setupContainerAskForRM.
/**
* Setup the request that will be sent to the RM for the container ask.
*
* @return the setup ResourceRequest to be sent to RM
*/
private ContainerRequest setupContainerAskForRM() {
// setup requirements for hosts
// using * as any host will do for the distributed shell app
// set the priority for the request
// TODO - what is the range for priority? how to decide?
Priority pri = Priority.newInstance(requestPriority);
// Set up resource type requirements
// For now, memory and CPU are supported so we set memory and cpu requirements
Resource capability = Resource.newInstance(containerMemory, containerVirtualCores);
ContainerRequest request = new ContainerRequest(capability, null, null, pri);
LOG.info("Requested container ask: " + request.toString());
return request;
}
use of org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest in project alluxio by Alluxio.
the class ContainerAllocator method requestContainers.
private void requestContainers() throws Exception {
String[] hosts;
boolean relaxLocality;
// YARN requires that priority for relaxed-locality requests is different from strict-locality.
Priority priority;
if (mPreferredHost != null) {
hosts = new String[] { mPreferredHost };
relaxLocality = mPreferredHost.equals("any");
priority = Priority.newInstance(100);
} else {
hosts = getPotentialWorkerHosts();
relaxLocality = true;
priority = Priority.newInstance(101);
}
int numContainersToRequest = mTargetNumContainers - mAllocatedContainers.size();
LOG.info("Requesting {} {} containers", numContainersToRequest, mContainerName);
if (hosts.length * mMaxContainersPerHost < numContainersToRequest) {
throw new RuntimeException(ExceptionMessage.YARN_NOT_ENOUGH_HOSTS.getMessage(numContainersToRequest, mContainerName, hosts.length));
}
ContainerRequest containerRequest = new ContainerRequest(mResource, hosts, null, /* any racks */
priority, relaxLocality);
LOG.info("Making {} resource request(s) for Alluxio {}s with cpu {} memory {}MB on hosts {}", numContainersToRequest, mContainerName, mResource.getVirtualCores(), mResource.getMemory(), hosts);
for (int i = 0; i < numContainersToRequest; i++) {
mRMClient.addContainerRequest(containerRequest);
}
}
Aggregations