Summary of resource management discussion

Eric W. Biederman ebiederm at xmission.com
Fri Mar 16 13:03:03 PDT 2007


Srivatsa Vaddagiri <vatsa at in.ibm.com> writes:

> On Thu, Mar 15, 2007 at 12:12:50PM -0700, Paul Menage wrote:
>> >Yes me too. But maybe to keep in simple in initial versions, we should
>> >avoid that optimisation and at the same time get statistics on duplicates?.
>> 
>> That's an implementation detail - we have more important points to
>> agree on right now ...
>
> yes :)
>
> Eric, did you have any opinion on this thread?

A little.  Sorry for the delay this is my low priority discussion to
follow.   Here comes a brain dump...

I do know of one case outside the context of containers/vservers where
I would have liked a nice multi-process memory limit.  HPC jobs doing
rdma have a bad habit of wanting to mlock all of memory and thus send
the memory allocator into a tail spin.  So the separate use would be nice.

I do suspect that we are likely to look at having a hierarchy for most
of the limits. For limits if we can stand the cost having a hierarchy
makes sense.  However I don't see the advantage of sharing the
hierarchy between different classes of resource groups.

I do know that implementing hierarchical counters are inefficient for
tracking resource consumption, and if we support deep hierarchies I
expect we will want to optimize that, in non-trivial ways.   At which
point the hierarchy will be composed of more that just a pointer to
it's parent.  We could easily end up with children taking a lease from
their parents saying you can have this many more before you look
upwards in the hierarchy again.  So with non-trivial hierarchy
information adding a pointer at a generic level doesn't seem to be
much of a help.

Further my gut feel is that in general we will want to limit all
resources on a per container basis.  At the same time I expect we will
want to limit other resources to select process groups or users even
farther inside of a container.  So a single hierarchy for multiple
resources seems a little questionable.

Which suggests that we want what I think was called multi-hierarchy
support.

I guess also with hierarchies and entering we probably need a limit
such that if you enter another resource group you can't do anything
except stay at the same level or descend into the hierarchy.  But
you can never remove your parent of that resource type.

With the whole shared subtree concept the mount namespace stores
things you can enter into in the mount namespace.  This overcomes
the difficulty of having to find a process who has the resources
you want when you want to enter someplace.  I think that concept
is a benefit of a filesystem interface.  Using mount and unmount to
pin magic subsystem state seems a little more natural to me then
having to do other filesystem manipulations.  Especially since
the mount namespace is reference counted so you can be certain
everything you have pinned will either remain visible to someone
or automatically disappear.   I don't like interfaces like
sysvipc that require manual destruction.  I do think there is some
sense in layout out a palette into the mount namespace we can enter
into.

There is another issue I'm not quite certain how to address.  To some
extent it makes sense to be able to compile out the resource
controllers ensuring their overhead goes away completely.  However for
the most part I think it makes sense to assume that when they are
compiled in we have an initial set of resource controllers that either
doesn't limit us at all or has very liberal limits that have the same
effect.  Dealing with the case of possibly limiting things when we
have the support compiled in does not seem to make a lot of sense to
me.

I hope that helps a little.  I'm still coming to terms with the issues
brought on by resource groups and controlling filesystems. 

Eric



More information about the Containers mailing list