Re: [PATCH] sched/fair: Rate limit calls to update_blocked_averages() for NOHZ

[Qais Yousef <qais.yousef@arm.com>]

On 01/25/21 14:23, Vincent Guittot wrote:
> On Fri, 22 Jan 2021 at 19:39, Qais Yousef <qais.yousef@arm.com> wrote:
> >
> > On 01/22/21 17:56, Vincent Guittot wrote:
> > > > ---
> > > >  kernel/sched/fair.c | 2 +-
> > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > >
> > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > > index 04a3ce20da67..fe2dc0024db5 100644
> > > > --- a/kernel/sched/fair.c
> > > > +++ b/kernel/sched/fair.c
> > > > @@ -8381,7 +8381,7 @@ static bool update_nohz_stats(struct rq *rq, bool force)
> > > >         if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
> > > >                 return false;
> > > >
> > > > -       if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
> > > > +       if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick + (HZ/20)))
> > >
> > > This condition is there to make sure to update blocked load at most
> > > once a tick in order to filter newly idle case otherwise the rate
> > > limit is already done by load balance interval
> > > This hard coded (HZ/20) looks really like an ugly hack
> >
> > This was meant as an RFC patch to discuss the problem really.
> >
> > Joel is seeing update_blocked_averages() taking ~100us. Half of it seems in
> > processing __update_blocked_fair() and the other half in sugov_update_shared().
> > So roughly 50us each. Note that each function is calling an iterator in
> 
> Can I assume that a freq change happens if sugov_update_shared() takes 50us ?
> which would mean that the update was useful at the end ?

I couldn't reproduce his problem on Juno. But I think it is not actually doing
any frequency update. sugov_update_shared() is rate limited by
sugov_should_update_freq(). Joel has a 1ms rate limit for schedutil in sysfs.
The function traces showed that it is continuously doing the full scan which
indicates that sugov_update_next_freq() is continuously bailing out at

	if else (sg_policy->next_freq == next_freq)
		return false;

> 
> > return. Correct me if my numbers are wrong Joel.
> >
> > Running on a little core on low frequency these numbers don't look too odd.
> > So I'm not seeing how we can speed these functions up.
> >
> > But since update_sg_lb_stats() will end up with multiple calls to
> > update_blocked_averages() in one go, this latency adds up quickly.
> >
> > One noticeable factor in Joel's system is the presence of a lot of cgroups.
> > Which is essentially what makes __update_blocked_fair() expensive, and it seems
> > to always return something has decayed so we end up with a call to
> > sugov_update_shared() in every call.
> >
> > I think we should limit the expensive call to update_blocked_averages() but
> 
> At the opposite, some will complain that block values  stay stall to
> high value and prevent any useful adjustment.
> 
> Also update_blocked average is already rate limited with idle and busy
> load_balance
> 
> Seems like the problem raised by Joel is the number of newly idle load balance

It could be. When Joel comments out the update_blocked_averages() or rate limit
it the big schedule delay disappears. Which is just an indication of where we
could do better. Either by making update_blocked_averages() faster, or control
how often we do it in a row. I thought the latter made more sense - though
implementation wise I'm not sure how we should do that.

We might actually help update_blocked_averages() being a bit faster by not
doing cpufreq_update_util() in every call and do it once in the last call to
update_blocked_averages(). Since it seemed the for_each_leaf_cfs_rq_safe() loop
in __update_blocked_fair() is expensive too, not sure if reducing the calls to
cpufreq_update_util() will be enough.

Thanks

--
Qais Yousef