[
https://issues.apache.org/jira/browse/MESOS-2713?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14536794#comment-14536794
]
Timothy Chen commented on MESOS-2713:
-------------------------------------
At the time we did the integration we just wanted to use existing statistics
code that's in Mesos, and I believe it was written to not assume cgroups is
present.
I'm not aware that using sysctl stats will be inaccurate, it should be roughly
the same as cgrups AFAIK.
> Docker resource usage
> ----------------------
>
> Key: MESOS-2713
> URL: https://issues.apache.org/jira/browse/MESOS-2713
> Project: Mesos
> Issue Type: Bug
> Components: containerization, docker, isolation
> Affects Versions: 0.22.1
> Reporter: Ian Babrou
>
> Looks like resource usage for docker containers on slaves is not very
> accurate (/monitor/statistics.json). For example, cpu usage is calculated by
> travesing process tree and summing up cpu times. Resulting numbers are not
> even close to real usage, CPU time can even decrease.
> What is the reason for this if you can use cgroup data directly? Reading
> cgroup location from pid of docker container is pretty straighforward.
> Another similar question: what is the reason to set isolation to posix
> instead of cgroups by default? Looks like it suffers from the same issues as
> docker containerizer (incorrect stats). More docs on this topic would be
> great.
> Posix isolation also leads to bigger CPU usage from mesos slave process
> (higher usage — posix isolation): http://i.imgur.com/jepk5m6.png
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
