[jira] [Commented] (MESOS-1113) Refactor cgroup interface in preparation for Systemd NWO.

["Joris Van Remoortere (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MESOS-1113?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14561547#comment-14561547 ] 

Joris Van Remoortere commented on MESOS-1113:
---------------------------------------------

We are running into this running systemd.
The symptom using 0.22.1 is that the cgroups directories "magically disappear" (systemd rebuilds the tree).
We will need to more tightly integrate with the systemd API to get this work properly, especially considering the long term changes mentioned (http://www.freedesktop.org/wiki/Software/systemd/ControlGroupInterface/).

I'd like to emphasize that some of our future work is going to be hampered by: 
{quote}For example cpuset or freezer are currently not exposed at all due to the broken inheritance semantics of the kernel logic{quote}

> Refactor cgroup interface in preparation for Systemd NWO.
> ---------------------------------------------------------
>
>                 Key: MESOS-1113
>                 URL: https://issues.apache.org/jira/browse/MESOS-1113
>             Project: Mesos
>          Issue Type: Bug
>          Components: containerization
>    Affects Versions: 0.19.0
>            Reporter: Timothy St. Clair
>            Assignee: Timothy St. Clair
>              Labels: cgroups
>
> In coming releases cgroups will no longer have it's own interface,  all interactions will go through systemd's DBUS interface:
> http://www.freedesktop.org/wiki/Software/systemd/ControlGroupInterface/
> This ticket is to track and allow the refactoring and migration that will be required in order to support.  
> ----- Original Message -----
> > From: "Lennart Poettering" <mz...@0pointer.de>
> > To: "Development discussions related to Fedora" <de...@lists.fedoraproject.org>
> > Cc: "Fedora Big Data SIG" <bi...@lists.fedoraproject.org>
> > Sent: Monday, March 17, 2014 8:18:37 PM
> > Subject: Re: Systemd & cgroups & NWO
> > 
> > Well, the nebulous choice of words is intended, since we don't want to
> > make specific promises on time-frames...
> > 
> > The APIs described (tersely) at the end of the wiki page describe the
> > status quo with systemd 211.
> > 
> > The "single-writer" cgroup tree stuff Tejun has been working on for the
> > kernel is now working on his machine, but it's not pushed upstream and
> > will take a while before it will hit Fedora.
> > 
> > At this point in time you hence still may create cgroups directly
> > yourself (but only if you follow the pax cgroup document), however, we
> > strongly encourage you to instead use scopes/slices to create them, as
> > discussed on the wiki page. This way the cgroups transition will be
> > abstracted away from you. You have control of a number of knobs that
> > systemd will expose for you, such as CPUShares=, BlockIOWeight= and so
> > on, but this is not complete, and primarily so because it's not clear
> > that those other properties will continue to exist the way they are in
> > the kernel. To read statistics data or to write knobs that systemd
> > doesn't cover you need to go directly to the cgroupfs. For that, simply
> > read /proc/self/cgroup to find out your own cgroup, and then operate on
> > that. However, as during the single-writer cgroup transition the kernel
> > interface how we set things up will change, be prepared that things
> > might break...
> > 
> > Lennart
> > 
> > --
> > Lennart Poettering, Red Hat



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)