[
https://issues.apache.org/jira/browse/YARN-1408?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14047755#comment-14047755
]
Wangda Tan commented on YARN-1408:
----------------------------------
Hi [~sunilg],
Thanks for updating the patch, overall approach LGTM, some comments,
1)
bq. I think we can have a new api in appSchedulingInfo to return list of
ResourceRequests (node local, rack local and any).
I would suggest to modify existing appSchedulingInfo.allocate to return list of
RRs. There existed outstanding resource decrement logic in allocate(), we can
simply add decremented RR to a list and return them. It looks more like
by-product of ASI.allocate to me.
2)
{code}
if (type.equals(NodeType.NODE_LOCAL)) {
list.add(nodeRequests.get(hostName));
}
{code}
It's better to clone RR instead of add ref to list. It works, but it's better
to set a #container correctly and prevent RR changed in ASI in the future.
3) TestCapacityScheduler:
It's good to have a test for FairScheduler here too. I think we can put the
test to org.apache.hadoop.yarn.server.resourcemanager.scheduler, and make it
parameterized for Fair/Capacity/FIFO.
Two minor comment for TestCapacityScheduler.
3.1
{code}
for (ResourceRequest request : requests) {
// Skip the OffRack and RackLocal resource requests.
if (request.getResourceName().equals(node.getRackName())
|| request.getResourceName().equals(ResourceRequest.ANY)) {
Assert.assertEquals(request.getNumContainers(), 1);
continue;
}
// Resource request must have added back in RM after preempt event
handling.
Assert.assertNotNull(app.getResourceRequest(request.getPriority(),
request.getResourceName()));
}
{code}
We can make it simpler to,
{code}
for (ResourceRequest request : requests) {
// Resource request must have added back in RM after preempt event
handling.
Assert.assertEquals(1, app.getResourceRequest(request.getPriority(),
request.getResourceName()).getNumContainers());
}
{code}
Because we added them back, there's no difference between node/rack/any.
3.2
{code}
// allocate container
List<Container> containers = am1.allocate(new ArrayList<ResourceRequest>(),
new ArrayList<ContainerId>()).getAllocatedContainers();
{code}
Should we wait for containers allocated in a while loop? This works now because
previous we called "rm1.waitForState(nm1, ...)". But it's better to wait
container allocated explictly
Thanks,
Wangda
> Preemption caused Invalid State Event: ACQUIRED at KILLED and caused a task
> timeout for 30mins
> ----------------------------------------------------------------------------------------------
>
> Key: YARN-1408
> URL: https://issues.apache.org/jira/browse/YARN-1408
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: resourcemanager
> Affects Versions: 2.2.0
> Reporter: Sunil G
> Assignee: Sunil G
> Attachments: Yarn-1408.1.patch, Yarn-1408.2.patch, Yarn-1408.3.patch,
> Yarn-1408.4.patch, Yarn-1408.5.patch, Yarn-1408.6.patch, Yarn-1408.patch
>
>
> Capacity preemption is enabled as follows.
> * yarn.resourcemanager.scheduler.monitor.enable= true ,
> *
> yarn.resourcemanager.scheduler.monitor.policies=org.apache.hadoop.yarn.server.resourcemanager.monitor.capacity.ProportionalCapacityPreemptionPolicy
> Queue = a,b
> Capacity of Queue A = 80%
> Capacity of Queue B = 20%
> Step 1: Assign a big jobA on queue a which uses full cluster capacity
> Step 2: Submitted a jobB to queue b which would use less than 20% of cluster
> capacity
> JobA task which uses queue b capcity is been preempted and killed.
> This caused below problem:
> 1. New Container has got allocated for jobA in Queue A as per node update
> from an NM.
> 2. This container has been preempted immediately as per preemption.
> Here ACQUIRED at KILLED Invalid State exception came when the next AM
> heartbeat reached RM.
> ERROR
> org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainerImpl:
> Can't handle this event at current state
> org.apache.hadoop.yarn.state.InvalidStateTransitonException: Invalid event:
> ACQUIRED at KILLED
> This also caused the Task to go for a timeout for 30minutes as this Container
> was already killed by preemption.
> attempt_1380289782418_0003_m_000000_0 Timed out after 1800 secs
--
This message was sent by Atlassian JIRA
(v6.2#6252)
