Re: [PATCH 0/7 v4] move update blocked load outside newidle_balance

[Vincent Guittot]

On Wed, 24 Feb 2021 at 18:41, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Wed, Feb 24, 2021 at 04:57:15PM +0100, Vincent Guittot wrote:
> > On Wed, 24 Feb 2021 at 16:54, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Wed, Feb 24, 2021 at 02:30:00PM +0100, Vincent Guittot wrote:
> > > > Joel reported long preempt and irq off sequence in newidle_balance because
> > > > of a large number of CPU cgroups in use and having to be updated. This
> > > > patchset moves the update outside newidle_imblance. This enables to early
> > > > abort during the updates in case of pending irq as an example.
> > > >
> > > > Instead of kicking a normal ILB that will wakes up CPU which is already
> > > > idle, patch 6 triggers the update of statistics in the idle thread of
> > > > the CPU before selecting and entering an idle state.
> > >
> > > I'm confused... update_blocked_averages(), which calls
> > > __update_blocked_fair(), which is the one doing the cgroup iteration
> > > thing, runs with rq->lock held, and thus will have IRQs disabled any
> > > which way around we turn this thing.
> > >
> > > Or is the problem that we called nohz_idle_balance(), which does
> > > update_nohz_stats() -> update_blocked_averages() for evey NOHZ cpu from
> > > newidle balance, such that we get NR_NOHZ_CPUS * NR_CGROUPS IRQ latency?
> > > Which is now reduced to just NR_CGROUPS ?
> >
> > Yes we can now abort between each cpu update
>
> OK, shall I add something like:
>
> This reduces the IRQ latency from O(nr_cgroups * nr_nohz_cpus) to
> O(nr_cgroups).
>
> To the changelog of patch #1 ?

Yes, good point. This will clarify the range of improvement