// This is to test container tokens are generated when the containers are
// acquired by the AM, not when the containers are allocated
@Test
public void testContainerTokenGeneratedOnPullRequest() throws Exception {
YarnConfiguration conf = new YarnConfiguration();
YarnAPIStorageFactory.setConfiguration(conf);
RMStorageFactory.setConfiguration(conf);
conf.setClass(YarnConfiguration.RM_SCHEDULER, CapacityScheduler.class, ResourceScheduler.class);
MockRM rm1 = new MockRM(conf);
try {
rm1.start();
MockNM nm1 = rm1.registerNode("127.0.0.1:1234", 8000);
RMApp app1 = rm1.submitApp(200);
MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
// request a container.
am1.allocate("127.0.0.1", 1024, 1, new ArrayList<ContainerId>());
ContainerId containerId2 = ContainerId.newInstance(am1.getApplicationAttemptId(), 2);
rm1.waitForState(nm1, containerId2, RMContainerState.ALLOCATED);
RMContainer container = rm1.getResourceScheduler().getRMContainer(containerId2);
// no container token is generated.
Assert.assertEquals(containerId2, container.getContainerId());
Assert.assertNull(container.getContainer().getContainerToken());
// acquire the container.
List<Container> containers =
am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>())
.getAllocatedContainers();
Assert.assertEquals(containerId2, containers.get(0).getId());
// container token is generated.
Assert.assertNotNull(containers.get(0).getContainerToken());
} finally {
rm1.stop();
}
}
@Test
public void testNormalContainerAllocationWhenDNSUnavailable() throws Exception {
YarnConfiguration conf = new YarnConfiguration();
YarnAPIStorageFactory.setConfiguration(conf);
RMStorageFactory.setConfiguration(conf);
MockRM rm1 = new MockRM(conf);
try {
rm1.start();
MockNM nm1 = rm1.registerNode("unknownhost:1234", 8000);
RMApp app1 = rm1.submitApp(200);
MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
// request a container.
am1.allocate("127.0.0.1", 1024, 1, new ArrayList<ContainerId>());
ContainerId containerId2 = ContainerId.newInstance(am1.getApplicationAttemptId(), 2);
rm1.waitForState(nm1, containerId2, RMContainerState.ALLOCATED);
// acquire the container.
SecurityUtilTestHelper.setTokenServiceUseIp(true);
List<Container> containers =
am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>())
.getAllocatedContainers();
// not able to fetch the container;
Assert.assertEquals(0, containers.size());
SecurityUtilTestHelper.setTokenServiceUseIp(false);
containers =
am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>())
.getAllocatedContainers();
// should be able to fetch the container;
Assert.assertEquals(1, containers.size());
} finally {
rm1.stop();
}
}
@Test
public void testPriorityWithPendingApplications() throws Exception {
Configuration conf = new Configuration();
conf.setClass(YarnConfiguration.RM_SCHEDULER, CapacityScheduler.class, ResourceScheduler.class);
// Set Max Application Priority as 10
conf.setInt(YarnConfiguration.MAX_CLUSTER_LEVEL_APPLICATION_PRIORITY, 10);
MockRM rm = new MockRM(conf);
rm.start();
Priority appPriority1 = Priority.newInstance(5);
MockNM nm1 = rm.registerNode("127.0.0.1:1234", 8 * GB);
RMApp app1 = rm.submitApp(1 * GB, appPriority1);
// kick the scheduler, 1 GB given to AM1, remaining 7GB on nm1
MockAM am1 = MockRM.launchAM(app1, rm, nm1);
am1.registerAppAttempt();
// kick the scheduler, 7 containers will be allocated for App1
List<Container> allocated1 = am1.allocateAndWaitForContainers("127.0.0.1", 7, 1 * GB, nm1);
Assert.assertEquals(7, allocated1.size());
Assert.assertEquals(1 * GB, allocated1.get(0).getResource().getMemory());
// check node report, 8 GB used (1 AM and 7 containers) and 0 GB available
SchedulerNodeReport report_nm1 = rm.getResourceScheduler().getNodeReport(nm1.getNodeId());
Assert.assertEquals(8 * GB, report_nm1.getUsedResource().getMemory());
Assert.assertEquals(0 * GB, report_nm1.getAvailableResource().getMemory());
// Submit the second app App2 with priority 7
Priority appPriority2 = Priority.newInstance(7);
RMApp app2 = rm.submitApp(1 * GB, appPriority2);
// Submit the third app App3 with priority 8
Priority appPriority3 = Priority.newInstance(8);
RMApp app3 = rm.submitApp(1 * GB, appPriority3);
// Submit the second app App4 with priority 6
Priority appPriority4 = Priority.newInstance(6);
RMApp app4 = rm.submitApp(1 * GB, appPriority4);
// Only one app can run as AM resource limit restricts it. Kill app1,
// If app3 (highest priority among rest) gets active, it indicates that
// priority is working with pendingApplications.
rm.killApp(app1.getApplicationId());
// kick the scheduler, app3 (high among pending) gets free space
MockAM am3 = MockRM.launchAM(app3, rm, nm1);
am3.registerAppAttempt();
// check node report, 1 GB used and 7 GB available
report_nm1 = rm.getResourceScheduler().getNodeReport(nm1.getNodeId());
Assert.assertEquals(1 * GB, report_nm1.getUsedResource().getMemory());
Assert.assertEquals(7 * GB, report_nm1.getAvailableResource().getMemory());
rm.stop();
}
@Test
public void testARRMResponseId() throws Exception {
MockNM nm1 = rm.registerNode("h1:1234", 5000);
RMApp app = rm.submitApp(2000);
// Trigger the scheduling so the AM gets 'launched'
nm1.nodeHeartbeat(true);
RMAppAttempt attempt = app.getCurrentAppAttempt();
MockAM am = rm.sendAMLaunched(attempt.getAppAttemptId());
am.registerAppAttempt();
AllocateRequest allocateRequest =
BuilderUtils.newAllocateRequest(attempt.getAppAttemptId(), 0, 0F, null, null);
AllocateResponse response = amService.allocate(allocateRequest);
Assert.assertEquals(1, response.getResponseId());
Assert.assertFalse(response.getReboot());
allocateRequest =
BuilderUtils.newAllocateRequest(
attempt.getAppAttemptId(), response.getResponseId(), 0F, null, null);
response = amService.allocate(allocateRequest);
Assert.assertEquals(2, response.getResponseId());
/* try resending */
response = amService.allocate(allocateRequest);
Assert.assertEquals(2, response.getResponseId());
/** try sending old request again * */
allocateRequest = BuilderUtils.newAllocateRequest(attempt.getAppAttemptId(), 0, 0F, null, null);
response = amService.allocate(allocateRequest);
Assert.assertTrue(response.getReboot());
}
@Test(timeout = 30000)
public void testExcessReservationThanNodeManagerCapacity() throws Exception {
YarnConfiguration conf = new YarnConfiguration();
YarnAPIStorageFactory.setConfiguration(conf);
RMStorageFactory.setConfiguration(conf);
MockRM rm = new MockRM(conf);
try {
rm.start();
// Register node1
MockNM nm1 = rm.registerNode("127.0.0.1:1234", 2 * GB, 4);
MockNM nm2 = rm.registerNode("127.0.0.1:2234", 3 * GB, 4);
nm1.nodeHeartbeat(true);
nm2.nodeHeartbeat(true);
// HOP :: Sleep to allow previous events to be processed
Thread.sleep(
conf.getInt(
YarnConfiguration.HOPS_PENDING_EVENTS_RETRIEVAL_PERIOD,
YarnConfiguration.DEFAULT_HOPS_PENDING_EVENTS_RETRIEVAL_PERIOD)
* 2);
// wait..
int waitCount = 20;
int size = rm.getRMContext().getActiveRMNodes().size();
while ((size = rm.getRMContext().getActiveRMNodes().size()) != 2 && waitCount-- > 0) {
LOG.info("Waiting for node managers to register : " + size);
Thread.sleep(100);
}
Assert.assertEquals(2, rm.getRMContext().getActiveRMNodes().size());
// Submit an application
RMApp app1 = rm.submitApp(128);
// kick the scheduling
nm1.nodeHeartbeat(true);
RMAppAttempt attempt1 = app1.getCurrentAppAttempt();
MockAM am1 = rm.sendAMLaunched(attempt1.getAppAttemptId(), nm1);
am1.registerAppAttempt();
LOG.info("sending container requests ");
am1.addRequests(new String[] {"*"}, 3 * GB, 1, 1);
AllocateResponse alloc1Response = am1.schedule(); // send the request
// kick the scheduler
nm1.nodeHeartbeat(true);
int waitCounter = 20;
LOG.info("heartbeating nm1");
while (alloc1Response.getAllocatedContainers().size() < 1 && waitCounter-- > 0) {
LOG.info("Waiting for containers to be created for app 1...");
Thread.sleep(500);
alloc1Response = am1.schedule();
}
LOG.info("received container : " + alloc1Response.getAllocatedContainers().size());
// No container should be allocated.
// Internally it should not been reserved.
Assert.assertTrue(alloc1Response.getAllocatedContainers().size() == 0);
LOG.info("heartbeating nm2");
waitCounter = 20;
nm2.nodeHeartbeat(true);
while (alloc1Response.getAllocatedContainers().size() < 1 && waitCounter-- > 0) {
LOG.info("Waiting for containers to be created for app 1...");
Thread.sleep(500);
alloc1Response = am1.schedule();
}
LOG.info("received container : " + alloc1Response.getAllocatedContainers().size());
Assert.assertTrue(alloc1Response.getAllocatedContainers().size() == 1);
} finally {
rm.stop();
}
}
@Test
public void testApplicationPriorityAllocation() throws Exception {
Configuration conf = new Configuration();
conf.setClass(YarnConfiguration.RM_SCHEDULER, CapacityScheduler.class, ResourceScheduler.class);
// Set Max Application Priority as 10
conf.setInt(YarnConfiguration.MAX_CLUSTER_LEVEL_APPLICATION_PRIORITY, 10);
MockRM rm = new MockRM(conf);
rm.start();
Priority appPriority1 = Priority.newInstance(5);
MockNM nm1 = rm.registerNode("127.0.0.1:1234", 16 * GB);
RMApp app1 = rm.submitApp(1 * GB, appPriority1);
// kick the scheduler, 1 GB given to AM1, remaining 15GB on nm1
MockAM am1 = MockRM.launchAM(app1, rm, nm1);
am1.registerAppAttempt();
// allocate 7 containers for App1
List<Container> allocated1 = am1.allocateAndWaitForContainers("127.0.0.1", 7, 2 * GB, nm1);
Assert.assertEquals(7, allocated1.size());
Assert.assertEquals(2 * GB, allocated1.get(0).getResource().getMemory());
// check node report, 15 GB used (1 AM and 7 containers) and 1 GB available
SchedulerNodeReport report_nm1 = rm.getResourceScheduler().getNodeReport(nm1.getNodeId());
Assert.assertEquals(15 * GB, report_nm1.getUsedResource().getMemory());
Assert.assertEquals(1 * GB, report_nm1.getAvailableResource().getMemory());
// Submit the second app App2 with priority 8 (Higher than App1)
Priority appPriority2 = Priority.newInstance(8);
RMApp app2 = rm.submitApp(1 * GB, appPriority2);
// kick the scheduler, 1 GB which was free is given to AM of App2
MockAM am2 = MockRM.launchAM(app2, rm, nm1);
am2.registerAppAttempt();
// check node report, 16 GB used and 0 GB available
report_nm1 = rm.getResourceScheduler().getNodeReport(nm1.getNodeId());
Assert.assertEquals(16 * GB, report_nm1.getUsedResource().getMemory());
Assert.assertEquals(0 * GB, report_nm1.getAvailableResource().getMemory());
// get scheduler
CapacityScheduler cs = (CapacityScheduler) rm.getResourceScheduler();
// get scheduler app
FiCaSchedulerApp schedulerAppAttempt =
cs.getSchedulerApplications().get(app1.getApplicationId()).getCurrentAppAttempt();
// kill 2 containers of App1 to free up some space
int counter = 0;
for (Container c : allocated1) {
if (++counter > 2) {
break;
}
cs.killContainer(schedulerAppAttempt.getRMContainer(c.getId()));
}
// check node report, 12 GB used and 4 GB available
report_nm1 = rm.getResourceScheduler().getNodeReport(nm1.getNodeId());
Assert.assertEquals(12 * GB, report_nm1.getUsedResource().getMemory());
Assert.assertEquals(4 * GB, report_nm1.getAvailableResource().getMemory());
// send updated request for App1
am1.allocate("127.0.0.1", 2 * GB, 10, new ArrayList<ContainerId>());
// kick the scheduler, since App2 priority is more than App1, it will get
// remaining cluster space.
List<Container> allocated2 = am2.allocateAndWaitForContainers("127.0.0.1", 2, 2 * GB, nm1);
// App2 has got 2 containers now.
Assert.assertEquals(2, allocated2.size());
// check node report, 16 GB used and 0 GB available
report_nm1 = rm.getResourceScheduler().getNodeReport(nm1.getNodeId());
Assert.assertEquals(16 * GB, report_nm1.getUsedResource().getMemory());
Assert.assertEquals(0 * GB, report_nm1.getAvailableResource().getMemory());
rm.stop();
}
@Test(timeout = 30000)
public void testNodeUpdate() throws Exception {
// set node -> label
mgr.addToCluserNodeLabels(ImmutableSet.of("x", "y", "z"));
// set mapping:
// h1 -> x
// h2 -> y
mgr.addLabelsToNode(ImmutableMap.of(NodeId.newInstance("h1", 0), toSet("x")));
mgr.addLabelsToNode(ImmutableMap.of(NodeId.newInstance("h2", 0), toSet("y")));
// inject node label manager
MockRM rm =
new MockRM(getConfigurationWithQueueLabels(conf)) {
@Override
public RMNodeLabelsManager createNodeLabelManager() {
return mgr;
}
};
rm.getRMContext().setNodeLabelManager(mgr);
rm.start();
MockNM nm1 = rm.registerNode("h1:1234", 8000);
MockNM nm2 = rm.registerNode("h2:1234", 8000);
MockNM nm3 = rm.registerNode("h3:1234", 8000);
ContainerId containerId;
// launch an app to queue a1 (label = x), and check all container will
// be allocated in h1
RMApp app1 = rm.submitApp(GB, "app", "user", null, "a");
MockAM am1 = MockRM.launchAndRegisterAM(app1, rm, nm3);
// request a container.
am1.allocate("*", GB, 1, new ArrayList<ContainerId>(), "x");
containerId = ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
Assert.assertTrue(rm.waitForState(nm1, containerId, RMContainerState.ALLOCATED, 10 * 1000));
// check used resource:
// queue-a used x=1G, ""=1G
checkUsedResource(rm, "a", 1024, "x");
checkUsedResource(rm, "a", 1024);
// change h1's label to z, container should be killed
mgr.replaceLabelsOnNode(ImmutableMap.of(NodeId.newInstance("h1", 0), toSet("z")));
Assert.assertTrue(rm.waitForState(nm1, containerId, RMContainerState.KILLED, 10 * 1000));
// check used resource:
// queue-a used x=0G, ""=1G ("" not changed)
checkUsedResource(rm, "a", 0, "x");
checkUsedResource(rm, "a", 1024);
// request a container with label = y
am1.allocate("*", GB, 1, new ArrayList<ContainerId>(), "y");
containerId = ContainerId.newContainerId(am1.getApplicationAttemptId(), 3);
Assert.assertTrue(rm.waitForState(nm2, containerId, RMContainerState.ALLOCATED, 10 * 1000));
// check used resource:
// queue-a used y=1G, ""=1G
checkUsedResource(rm, "a", 1024, "y");
checkUsedResource(rm, "a", 1024);
// change h2's label to no label, container should be killed
mgr.replaceLabelsOnNode(
ImmutableMap.of(NodeId.newInstance("h2", 0), CommonNodeLabelsManager.EMPTY_STRING_SET));
Assert.assertTrue(rm.waitForState(nm1, containerId, RMContainerState.KILLED, 10 * 1000));
// check used resource:
// queue-a used x=0G, y=0G, ""=1G ("" not changed)
checkUsedResource(rm, "a", 0, "x");
checkUsedResource(rm, "a", 0, "y");
checkUsedResource(rm, "a", 1024);
containerId = ContainerId.newContainerId(am1.getApplicationAttemptId(), 1);
// change h3's label to z, AM container should be killed
mgr.replaceLabelsOnNode(ImmutableMap.of(NodeId.newInstance("h3", 0), toSet("z")));
Assert.assertTrue(rm.waitForState(nm1, containerId, RMContainerState.KILLED, 10 * 1000));
// check used resource:
// queue-a used x=0G, y=0G, ""=1G ("" not changed)
checkUsedResource(rm, "a", 0, "x");
checkUsedResource(rm, "a", 0, "y");
checkUsedResource(rm, "a", 0);
rm.close();
}
@Test
public void testAMRMUnusableNodes() throws Exception {
MockNM nm1 = rm.registerNode("h1:1234", 5000);
MockNM nm2 = rm.registerNode("h2:1234", 5000);
MockNM nm3 = rm.registerNode("h3:1234", 5000);
MockNM nm4 = rm.registerNode("h4:1234", 5000);
RMApp app1 = rm.submitApp(2000);
// Trigger the scheduling so the AM gets 'launched' on nm1
nm1.nodeHeartbeat(true);
RMAppAttempt attempt1 = app1.getCurrentAppAttempt();
MockAM am1 = rm.sendAMLaunched(attempt1.getAppAttemptId());
// register AM returns no unusable node
am1.registerAppAttempt();
// allocate request returns no updated node
AllocateRequest allocateRequest1 =
BuilderUtils.newAllocateRequest(attempt1.getAppAttemptId(), 0, 0F, null, null);
AMResponse response1 = amService.allocate(allocateRequest1).getAMResponse();
List<NodeReport> updatedNodes = response1.getUpdatedNodes();
Assert.assertEquals(0, updatedNodes.size());
syncNodeHeartbeat(nm4, false);
// allocate request returns updated node
allocateRequest1 =
BuilderUtils.newAllocateRequest(
attempt1.getAppAttemptId(), response1.getResponseId(), 0F, null, null);
response1 = amService.allocate(allocateRequest1).getAMResponse();
updatedNodes = response1.getUpdatedNodes();
Assert.assertEquals(1, updatedNodes.size());
NodeReport nr = updatedNodes.iterator().next();
Assert.assertEquals(nm4.getNodeId(), nr.getNodeId());
Assert.assertEquals(NodeState.UNHEALTHY, nr.getNodeState());
// resending the allocate request returns the same result
response1 = amService.allocate(allocateRequest1).getAMResponse();
updatedNodes = response1.getUpdatedNodes();
Assert.assertEquals(1, updatedNodes.size());
nr = updatedNodes.iterator().next();
Assert.assertEquals(nm4.getNodeId(), nr.getNodeId());
Assert.assertEquals(NodeState.UNHEALTHY, nr.getNodeState());
syncNodeLost(nm3);
// subsequent allocate request returns delta
allocateRequest1 =
BuilderUtils.newAllocateRequest(
attempt1.getAppAttemptId(), response1.getResponseId(), 0F, null, null);
response1 = amService.allocate(allocateRequest1).getAMResponse();
updatedNodes = response1.getUpdatedNodes();
Assert.assertEquals(1, updatedNodes.size());
nr = updatedNodes.iterator().next();
Assert.assertEquals(nm3.getNodeId(), nr.getNodeId());
Assert.assertEquals(NodeState.LOST, nr.getNodeState());
// registering another AM gives it the complete failed list
RMApp app2 = rm.submitApp(2000);
// Trigger nm2 heartbeat so that AM gets launched on it
nm2.nodeHeartbeat(true);
RMAppAttempt attempt2 = app2.getCurrentAppAttempt();
MockAM am2 = rm.sendAMLaunched(attempt2.getAppAttemptId());
// register AM returns all unusable nodes
am2.registerAppAttempt();
// allocate request returns no updated node
AllocateRequest allocateRequest2 =
BuilderUtils.newAllocateRequest(attempt2.getAppAttemptId(), 0, 0F, null, null);
AMResponse response2 = amService.allocate(allocateRequest2).getAMResponse();
updatedNodes = response2.getUpdatedNodes();
Assert.assertEquals(0, updatedNodes.size());
syncNodeHeartbeat(nm4, true);
// both AM's should get delta updated nodes
allocateRequest1 =
BuilderUtils.newAllocateRequest(
attempt1.getAppAttemptId(), response1.getResponseId(), 0F, null, null);
response1 = amService.allocate(allocateRequest1).getAMResponse();
updatedNodes = response1.getUpdatedNodes();
Assert.assertEquals(1, updatedNodes.size());
nr = updatedNodes.iterator().next();
Assert.assertEquals(nm4.getNodeId(), nr.getNodeId());
Assert.assertEquals(NodeState.RUNNING, nr.getNodeState());
allocateRequest2 =
BuilderUtils.newAllocateRequest(
attempt2.getAppAttemptId(), response2.getResponseId(), 0F, null, null);
response2 = amService.allocate(allocateRequest2).getAMResponse();
updatedNodes = response2.getUpdatedNodes();
Assert.assertEquals(1, updatedNodes.size());
nr = updatedNodes.iterator().next();
Assert.assertEquals(nm4.getNodeId(), nr.getNodeId());
Assert.assertEquals(NodeState.RUNNING, nr.getNodeState());
// subsequent allocate calls should return no updated nodes
allocateRequest2 =
BuilderUtils.newAllocateRequest(
attempt2.getAppAttemptId(), response2.getResponseId(), 0F, null, null);
response2 = amService.allocate(allocateRequest2).getAMResponse();
updatedNodes = response2.getUpdatedNodes();
Assert.assertEquals(0, updatedNodes.size());
// how to do the above for LOST node
}