Re: [PATCH, v3 2/2] cgroups: introduce timer slack subsystem

[Matt Helsley <matthltc@us.ibm.com>]

On Fri, Feb 04, 2011 at 09:27:55AM -0800, Jacob Pan wrote:
> On Fri, 4 Feb 2011 15:34:39 +0200
> "Kirill A. Shutemov" <kirill@shutemov.name> wrote:
> 
> > On Thu, Feb 03, 2011 at 11:57:43AM -0800, Jacob Pan wrote:
> > > On Thu, 3 Feb 2011 10:12:51 -0800
> > > Paul Menage <menage@google.com> wrote:
> > > 
> > > > On Thu, Feb 3, 2011 at 9:51 AM, Jacob Pan
> > > > <jacob.jun.pan@linux.intel.com> wrote:
> > > > >
> > > > > I think this logic defeats the purpose of having timer_slack
> > > > > subsystem in the first place. IMHO, the original intention was
> > > > > to have grouping effect of tasks in the cgroup.
> > > > 
> > > > You can get the semantics you want by just setting min_slack_ns =
> > > > max_slack_ns.
> > > > 
> > > true. it will just make set fail when min = max. it is awkward and
> > > counter intuitive when you want to change the group timer_slack. you
> > > will have to move both min and max to clamp the value, where set
> > > function can not be used.
> > 
> > Interface is very similar to /sys/devices/system/cpu/cpuX/cpufreq.
> > I think it's sane. If you want some extention, you can do it with
> > userspace helper.
> > 
> I don't disagree the current interface is usable. Just less intuitive.
> The situation is different for cpufreq, where you
> don't the situation of adding new entries to be adjusted in the
> existing max-min range.

We can probably eliminate the upper end of the range for now -- I don't
see any practical reason for it and we can always add it later if I'm
being too short-sighted. :)

We could also probably eliminate the "set" interface and rely solely on
prctl() for that. The only thing needed is a "minimum" timer slack
applied for the cgroup. We could apply that minimum like the current
code does _or_ via the same method as option #2 which you presented below.

>
> 
> > > In addition, when a parent changes min = max, I don't see the
> > > current code enforce new settings on the children. Am i missing
> > > something?
> > 
> > I've missed it. I'll fix.

Sounds good. Allowing child cgroups also allow for nesting of containers or
imposing limits to timer slack on users and all of their containers.

> > 
> > > In my use case, i want to put some apps into a managed group where
> > > relaxed slack value is used, but when time comes to move the app
> > > out of that cgroup, we would like to resore the original timer
> > > slack. I see having a current value per cgroup can be useful if we
> > > let timer code pick whether to use task slack value or the cgroup
> > > slack value. Or we have to cache the old value per task
> > 
> > What's mean "original timer slack" if you are free to move a task
> > between a lot of cgroups and process itself free to change it anytime?
> > 
> 
> I need to manage tasks by a management software instead of letting the
> task change timer_slack by itself. The goal is to make management
> transparent and no modifications to the existing apps. Therefore, it is

Until those apps learn that there's a subsystem they can manipulate
too ;).

> desirable to automatically enforce timer_slack when the apps are in the
> cgroup while automatically restore it when it is no longer under cgroup
> management.
>
> So the "original timer slack" can be the default 50us or whatever value
> chosen by the task itself. But the app itself should not care or even be
> aware of which cgroup it is in.

I don't think anyone is arguing it should.

> 
> So here are two optoins i can think of
> 1. add a new variable called cg_timer_slack_ns to struct task_struct{}
> cg_timer_slack_ns will be set by cgroup timer_slack subsystem, then we
> can retain the original per task value in timer_slack_ns.
> timer code will pick max(cg_timer_slack_ns, timer_slack_ns) if
> cg_timer_slack_ns is set.
> 
> 2. leave task_struct unchanged, add a current_timer_slack to the
> cgroup. timer_slack cgroup does not modify per task timer_slack_ns.
> similar to option #1, let timer code pick the timer_slack to use based
> on whether the task is in timer_slack cgroup.

I really like #2 and strongly dislike #1 because cgroup subsystem
information should stay out of the task struct as much as possible.

Cheers,
	-Matt Helsley