[
https://issues.apache.org/jira/browse/MESOS-2767?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14565232#comment-14565232
]
Joris Van Remoortere commented on MESOS-2767:
---------------------------------------------
After talking with [~idownes] I want to clarify the impact of this change:
Once the usage() call is switched from {{mesos::internal::usage}} to using the
cgroups there are at least 2 big changes:
1) Cgroups will report total accumulated statistics for all the pids that were
in the group. This means it will account for ephemeral pids. The
{{mesos::internal::usage}} version only reports the amount for the current
active pid. This means the new cpu usage reports will be more consistent, but
also different.
2) Cgroups calculates the memory usage differently from
{{mesos::internal::usage}}. [~idownes] has a better understanding of this, and
has left some comments in the relevant sections of code.
The consequence of the above is that users of these usage metrics should be
careful when upgrading to this potential change, as it may trigger monitoring
alerts. Any tooling around these may need to be tuned to adjust for the
accounting changes.
If we continue with this change, we will need to notify the dev list, as well
as make it clear in the release notes.
> Gather usage of containers from DockerContainerizer via cgroups
> ---------------------------------------------------------------
>
> Key: MESOS-2767
> URL: https://issues.apache.org/jira/browse/MESOS-2767
> Project: Mesos
> Issue Type: Improvement
> Components: docker
> Reporter: Timothy Chen
>
> Currently the DockerContaierizer in usage() is calling mesos::internal::usage
> which is gathering usage by calling sysctl and walking through every process
> in the system.
> We should directly gather statitstics by looking at cgroups, simliar of how
> MesosContainerizer or cadvisor works.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
