[Devel] Re: [ckrm-tech] [PATCH] BC: resource beancounters (v4) (added user memory)

[Rohit Seth]

On Tue, 2006-09-12 at 16:54 -0700, Chandra Seetharaman wrote:
> On Mon, 2006-09-11 at 16:58 -0700, Rohit Seth wrote:
> > On Mon, 2006-09-11 at 12:42 -0700, Chandra Seetharaman wrote:
> > > On Mon, 2006-09-11 at 12:10 -0700, Rohit Seth wrote:
> > > > On Mon, 2006-09-11 at 11:25 -0700, Chandra Seetharaman wrote:
> > 
> > > > > There could be a default container which doesn't have any guarantee or
> > > > > limit. 
> > > > 
> > > > First, I think it is critical that we allow processes to run outside of
> > > > any container (unless we know for sure that the penalty of running a
> > > > process inside a container is very very minimal).
> > > 
> > > When I meant a default container I meant a default "resource group". In
> > > case of container that would be the default environment. I do not see
> > > any additional overhead associated with it, it is only associated with
> > > how resource are allocated/accounted.
> > > 
> > 
> > There should be some cost when you do atomic inc/dec accounting and
> > locks for add/remove resources from any container (including default
> > resource group). No?
> 
> yes, it would be there, but is not heavy, IMO.

I think anything greater than 1% could be a concern for people who are
not very interested in containers but would be forced to live with them.

> > 
> > > > 
> > > > And anything running outside a container should be limited by default
> > > > Linux settings.
> > > 
> > > note that the resource available to the default RG will be (total system
> > > resource - allocated to RGs).
> > 
> > I think it will be preferable to not change the existing behavior for
> > applications that are running outside any container (in your case
> > default resource group).
> 
> hmm, when you provide QoS for a set of apps, you will affect (the
> resource availability of) other apps. I don't see any way around it. Any
> ideas ?

When I say, existing behavior, I mean not getting impacted by some
artificial limits that are imposed by container subsystem.  IOW, if a
sysadmin is okay to have certain apps running outside of container then
he is basically forgoing any QoS for any container on that system.

>  
> > 
> > > > 
> > > > > When you create containers and assign guarantees to each of them
> > > > > make sure that you leave some amount of resource unassigned. 
> > > >                            ^^^^^ This will force the "default" container
> > > > with limits (indirectly). IMO, the whole guarantee feature gets defeated
> > > 
> > > You _will_ have limits for the default RG even if we don't have
> > > guarantees.
> > > 
> > > > the moment you bring in this fuzziness.
> > > 
> > > Not really. 
> > >  - Each RG will have a guarantee and limit of each resource. 
> > >  - default RG will have (system resource - sum of guarantees)
> > >  - Every RG will be guaranteed some amount of resource to provide QoS
> > >  - Every RG will be limited at "limit" to prevent DoS attacks.
> > >  - Whoever doesn't care either of those set them to don't care values.
> > > 
> > 
> > For the cases that put this don't care, do you depend on existing
> > reclaim algorithm (for memory) in kernel?
> 
> Yes.

So one container with these don't care condition(s) can turn the whole
guarantee thing bad.  Because existing kernel reclaimer does not know
about memory commitments to other containers.  Right?

> > 
> > > > 
> > > > > That
> > > > > unassigned resources can be used by the default container or can be used
> > > > > by containers that want more than their guarantee (and less than their
> > > > > limit). This is how CKRM/RG handles this issue.
> > > > > 
> > > > >  
> > > > 
> > > > It seems that a single notion of limit should suffice, and that limit
> > > > should more be treated as something beyond which that resource
> > > > consumption in the container will be throttled/not_allowed.
> > > 
> > > As I stated in an earlier email "Limit only" approach can prevent a
> > > system from DoS attacks (and also fits the container model nicely),
> > > whereas to provide QoS one would need guarantee.
> > > 
> > > Without guarantee, a RG that the admin cares about can starve if
> > > all/most of the other RGs consume upto their limits.
> > > 
> > > > 
> > 
> > If the limits are set appropriately so that containers total memory
> > consumption does not exceed the system memory then there shouldn't be
> > any QoS issue (to whatever extent it is applicable for specific
> > scenario).
> 
> Then you will not be work-conserving (IOW over-committing), which is one
> of the main advantage of this type of feature.
> 

If for the systems where QoS is important, not over-committing will be
fine (at least to start with).

-rohit