[
https://issues.apache.org/jira/browse/YARN-7872?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16407448#comment-16407448
]
Wangda Tan commented on YARN-7872:
----------------------------------
[~yqwang],
Like [~jlowe] mentioned, currently we don't have an option to say "I don't care
about node partition". Non-exclusive node partition has different semantics.
The reason why we don't have it is, we need to calculate total pending resource
to each partition to make sure preemption can work properly.
Existing the patch is not backward compatible: it breaks behavior of an app
requests locality + node partition. Before your patch, the behavior is, if
requested locality is under requested partition, it can be allocated, otherwise
it will keep in pending state. After your patch, requested partition will be
silently ignored, which is not ideal. And it breaks how we calculate pending
resource of each partition.
Unfortunately I cannot think of a way to make it work. The good thing is, In
practice it may not be that bad:
* For traditional batch jobs, like MR/Spark, it will downgrade to
rack/offswitch read.
* For hard locality (relax locality == false), this causes task cannot get
allocated. This happens under misconfiguration, which is not normal case as
well.
Thoughts?
> labeled node cannot be used to satisfy locality specified request
> -----------------------------------------------------------------
>
> Key: YARN-7872
> URL: https://issues.apache.org/jira/browse/YARN-7872
> Project: Hadoop YARN
> Issue Type: Bug
> Components: capacity scheduler, capacityscheduler, resourcemanager
> Affects Versions: 2.7.2
> Reporter: Yuqi Wang
> Assignee: Yuqi Wang
> Priority: Blocker
> Fix For: 2.7.2
>
> Attachments: YARN-7872-branch-2.7.2.001.patch
>
>
> *Issue summary:*
> labeled node (i.e. node with 'not empty' node label) cannot be used to
> satisfy locality specified request (i.e. container request with 'not ANY'
> resource name and the relax locality is false).
>
> *For example:*
> The node with available resource:
> [Resource: [MemoryMB: [100] CpuNumber: [12]] {color:#14892c}NodeLabel:
> [persistent]{color} {color:#f79232}HostName: \{SRG}{color} RackName:
> \{/default-rack}]
> The container request:
> [Priority: [1] Resource: [MemoryMB: [1] CpuNumber: [1]]
> {color:#14892c}NodeLabel: [null]{color} {color:#f79232}HostNames:
> \{SRG}{color} RackNames: {} {color:#59afe1}RelaxLocality: [false]{color}]
> Current RM capacity scheduler's behavior is that (at least for version 2.7
> and 2.8), the node cannot allocate container for the request, because the
> node label is not matched when the leaf queue assign container.
>
> *Possible solution:*
> However, node locality and node label should be two orthogonal dimensions to
> select candidate nodes for container request. And the node label matching
> should only be executed for container request with ANY resource name, since
> only this kind of container request is allowed to have 'not empty' node label.
> So, for container request with 'not ANY' resource name (so, we clearly know
> it should not have node label), we should use the requested resource name to
> match with the node instead of using the requested node label to match with
> the node. And this resource name matching should be safe, since the node
> whose node label is not accessible for the queue will not be sent to the leaf
> queue.
>
> *Discussion:*
> Attachment is the fix according to this principle, please help to review.
> Without it, we cannot use locality to request container within these labeled
> nodes.
> If the fix is acceptable, we should also recheck whether the same issue
> happens in trunk and other hadoop versions.
> If not acceptable (i.e. the current behavior is by designed), so, how can we
> use locality to request container within these labeled nodes?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
