[
https://issues.apache.org/jira/browse/YARN-5301?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15944419#comment-15944419
]
Miklos Szegedi commented on YARN-5301:
--------------------------------------
I was able to reproduce this with Ubuntu, kernel 4.4. The main issue is that
Yarn currently supports two distinct options. We either scan for existing mount
points or create them. In case of newer OS versions cgroups is premounted,
usually cpu with cpuacct. If we want to remount it, we need to know whether it
was mounted together with other subsystems or not. This check is missing. To
make things more complicated a system may have some subsystems pre-mounted but
others not.
I suggest the following:
1. Check for existing mount points even if the mount option is enabled
2. If we select mount but there is a preexisting mount point with the current
subsystem, then mount it with the accompanying subsystems correctly depending
on the current system.
Since [~sandflee] released this jira, I started to work on a patch.
> NM mount cpu cgroups failed on some system
> ------------------------------------------
>
> Key: YARN-5301
> URL: https://issues.apache.org/jira/browse/YARN-5301
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: sandflee
>
> on ubuntu with linux kernel 3.19, , NM start failed if enable auto mount
> cgroup. try command:
> ./bin/container-executor --mount-cgroups yarn-hadoop cpu=/cgroup/cpu fail
> ./bin/container-executor --mount-cgroups yarn-hadoop cpu,cpuacct=/cgroup/cpu
> succ
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
