[
https://issues.apache.org/jira/browse/YARN-4599?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15129300#comment-15129300
]
MENG DING commented on YARN-4599:
---------------------------------
I have a question regarding when OOM controller will kick in. Is it true that:
* If memory.swappiness is set to 0, then OOM controller will kick in when
{{memory.limit_in_bytes}} (hard limit) is reached.
* If memory.swappiness is not set to 0, then OOM controller will kick in only
when all available swap in the system is used up, as
{{memory.memsw.limit_in_bytes}} is not a configurable parameter in YARN right
now?
Looking at this link, I am a little bit confused:
https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/6/html/Resource_Management_Guide/sec-memory.html
bq. Consider the following example: setting memory.limit_in_bytes = 2G and
memory.memsw.limit_in_bytes = 4G for a certain cgroup will allow processes in
that cgroup to allocate 2 GB of memory and, once exhausted, allocate another 2
GB of swap only. The memory.memsw.limit_in_bytes parameter represents the sum
of memory and swap. Processes in a cgroup that does not have the
memory.memsw.limit_in_bytes parameter set can potentially use up all the
available swap (after exhausting the set memory limitation) and trigger an Out
Of Memory situation caused by the lack of available swap.
bq. *memory.swappiness* Note that a value of 0 does not prevent process memory
being swapped out; swap out might still happen when there is a shortage of
system memory because the global virtual memory management logic does not read
the cgroup value.
> Set OOM control for memory cgroups
> ----------------------------------
>
> Key: YARN-4599
> URL: https://issues.apache.org/jira/browse/YARN-4599
> Project: Hadoop YARN
> Issue Type: Bug
> Components: nodemanager
> Affects Versions: 2.9.0
> Reporter: Karthik Kambatla
> Assignee: Karthik Kambatla
>
> YARN-1856 adds memory cgroups enforcing support. We should also explicitly
> set OOM control so that containers are not killed as soon as they go over
> their usage. Today, one could set the swappiness to control this, but
> clusters with swap turned off exist.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
