Re: [PATCH RFC 0/3] sched/fair: cpu time reserves for cgroups

[Konstantin Khlebnikov]

On 16.05.2016 14:18, Peter Zijlstra wrote:
> On Mon, May 16, 2016 at 12:36:19PM +0300, Konstantin Khlebnikov wrote:
> > This feature allows to change cpu cgroup weight for a limited time.
> >
> > Cgroup interface:
> > cpu.cfs_reserve_us      - reserved time for each cpu.cfs_period_us
> > cpu.cfs_reserve_shares  - group weight during reserved time
> >
> > While cfs group consumes reserved cpu time it has different weight,
> > thus it gets different vruntime penalty for that execution.
> >
> >          ^ weight
> >          |
> >          |
> > reserve |
> > shares  -------*
> >          |      |
> >          |      |
> >          |      |
> > shares  -      *-----------------*
> >          |                        |
> >          |                        |
> >          0------|-----------------|-----------> time
> >                 reserve           quota
> >
> > Reserve can work as a "low limit": boost weight for "guaranteed" time,
> > and as a "high limit": give normal weight for a limited time and allow
> > utilize cpu when nobody else needs it.
>
> You forgot to explain why I should care about this.

As I told this works as low-limit or high-limit which allow to
control cpu time distribution without hard limits and throttling.

Present quota/hard limit has well known problems when it throttle task
inside kernel where it holds mutexes. Also it's too strict and doesn't
allow utilization of unused cpu time.

--
Konstantin