YARN-4230. RM crashes with NPE when increasing container resource if there is no headroom left. Contributed by Meng Ding

This commit is contained in:
Jian He 2015-10-12 11:51:33 -07:00
parent e617cf6dd1
commit 9849c8b386
3 changed files with 77 additions and 2 deletions

View File

@ -926,6 +926,9 @@ Release 2.8.0 - UNRELEASED
YARN-4201. AMBlacklist does not work for minicluster. (Jun Gong via zxu)
YARN-4230. RM crashes with NPE when increasing container resource if there is no headroom left.
(Meng Ding via jianhe)
Release 2.7.2 - UNRELEASED
INCOMPATIBLE CHANGES

View File

@ -323,8 +323,8 @@ public CSAssignment assignContainers(Resource clusterResource,
}
}
// We already allocated something
if (!assigned.getSkipped()) {
// We may have allocated something
if (assigned != null && !assigned.getSkipped()) {
break;
}
}

View File

@ -284,6 +284,78 @@ public RMNodeLabelsManager createNodeLabelManager() {
rm1.close();
}
@Test
public void testIncreaseRequestWithNoHeadroomLeft() throws Exception {
/**
* Application has two containers running, try to increase one of them, the
* requested amount exceeds user's headroom for the queue.
*/
MockRM rm1 = new MockRM() {
@Override
public RMNodeLabelsManager createNodeLabelManager() {
return mgr;
}
};
rm1.start();
MockNM nm1 = rm1.registerNode("h1:1234", 8 * GB);
// app1 -> a1
RMApp app1 = rm1.submitApp(1 * GB, "app", "user", null, "default");
MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
FiCaSchedulerApp app = getFiCaSchedulerApp(rm1, app1.getApplicationId());
// Allocate 1 container
am1.allocate(
Arrays.asList(ResourceRequest.newInstance(Priority.newInstance(1), "*",
Resources.createResource(2 * GB), 1)),
null);
ContainerId containerId2 =
ContainerId.newContainerId(am1.getApplicationAttemptId(), 2);
Assert.assertTrue(rm1.waitForState(nm1, containerId2,
RMContainerState.ALLOCATED, 10 * 1000));
// Acquire them, and NM report RUNNING
am1.allocate(null, null);
sentRMContainerLaunched(rm1, containerId2);
// am1 asks to change container2 from 2GB to 8GB, which will exceed user
// limit
am1.sendContainerResizingRequest(Arrays.asList(
ContainerResourceChangeRequest
.newInstance(containerId2, Resources.createResource(8 * GB))),
null);
checkPendingResource(rm1, "default", 6 * GB, null);
Assert.assertEquals(6 * GB,
app.getAppAttemptResourceUsage().getPending().getMemory());
// NM1 do 1 heartbeats
CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
RMContainer rmContainer1 = app.getLiveContainersMap().get(containerId2);
/* Check reservation statuses */
// Increase request should *NOT* be reserved as it exceeds user limit
Assert.assertFalse(rmContainer1.hasIncreaseReservation());
Assert.assertTrue(app.getReservedContainers().isEmpty());
Assert.assertNull(cs.getNode(nm1.getNodeId()).getReservedContainer());
// Pending resource will not be changed since it's not satisfied
checkPendingResource(rm1, "default", 6 * GB, null);
Assert.assertEquals(6 * GB,
app.getAppAttemptResourceUsage().getPending().getMemory());
// Queue/user/application's usage will *NOT* be updated
checkUsedResource(rm1, "default", 3 * GB, null);
Assert.assertEquals(3 * GB, ((LeafQueue) cs.getQueue("default"))
.getUser("user").getUsed().getMemory());
Assert.assertEquals(3 * GB,
app.getAppAttemptResourceUsage().getUsed().getMemory());
Assert.assertEquals(0 * GB,
app.getAppAttemptResourceUsage().getReserved().getMemory());
rm1.close();
}
@Test
public void testExcessiveReservationWhenCancelIncreaseRequest()
throws Exception {