[
https://issues.apache.org/jira/browse/YARN-8930?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16659761#comment-16659761
]
Haibo Chen commented on YARN-8930:
----------------------------------
The patch allows the polling-based memory check to kick in, when the
strict-memory-enforcement with cgroup and elastic memory control are both
enabled.
When elastic memory control is disabled and cgroup-based strict memory
enforcement is enabled, the oom-killer is enabled on individual container
cgroup,
the kernel would automatically invoke the oom killer to handle any container
that runs out of memory limit, so check the under_oom status is unnecessary.
> CGroup-based strict container memory enforcement does not work with
> CGroupElasticMemoryController
> -------------------------------------------------------------------------------------------------
>
> Key: YARN-8930
> URL: https://issues.apache.org/jira/browse/YARN-8930
> Project: Hadoop YARN
> Issue Type: Bug
> Components: nodemanager
> Affects Versions: 3.2.0
> Reporter: Haibo Chen
> Assignee: Haibo Chen
> Priority: Major
> Attachments: YARN-8930.00.patch
>
>
> When yarn.nodemanger.resource.memory.enforced is set to true with memory
> cgroup turned on, (aka strict memory enforcement), containers monitor relies
> on the under_oom status read from the container cgroup's memory.oom_control
> file.
> However, when the root yarn container cgroup is under oom (e.g. when the node
> is overallocating iteself), the under_oom status is set for all yarn
> containers regardless of whether each individual container has run over its
> memory limit.
> What essentially happens is that whenever the root cgroup is under oom, all
> yarn containers are killed.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
