Re: [PATCH 3/3] sched: Spare idle load balancing on nohz_full CPUs

[Mike Galbraith <efault@gmx.de>]

On Tue, 2017-06-20 at 13:42 -0400, Rik van Riel wrote:
> On Mon, 2017-06-19 at 04:12 +0200, Frederic Weisbecker wrote:
> > Although idle load balancing obviously only concern idle CPUs, it can
> > be a disturbance on a busy nohz_full CPU. Indeed a CPU can only get
> > rid
> > of an idle load balancing duty once a tick fires while it runs a task
> > and this can take a while in a nohz_full CPU.
> > 
> > We could fix that and escape the idle load balancing duty from the
> > very
> > idle exit path but that would bring unecessary overhead. Lets just
> > not
> > bother and leave that job to housekeeping CPUs (those outside
> > nohz_full
> > range). The nohz_full CPUs simply don't want any disturbance.
> > 
> > Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > Cc: Ingo Molnar <mingo@kernel.org>
> > Cc: Rik van Riel <riel@redhat.com>
> > Cc: Peter Zijlstra <peterz@infradead.org>
> > ---
> >  kernel/sched/fair.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> > 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index d711093..cfca960 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -8659,6 +8659,10 @@ void nohz_balance_enter_idle(int cpu)
> >  	if (!cpu_active(cpu))
> >  		return;
> >  
> > +	/* Spare idle load balancing on CPUs that don't want to be
> > disturbed */
> > +	if (!is_housekeeping_cpu(cpu))
> > +		return;
> > +
> >  	if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
> >  		return;
> 
> I am not entirely convinced on this one.
> 
> Doesn't the if (on_null_domain(cpu_rq(cpu)) test
> a few lines down take care of this already?
> 
> Do we want nohz_full to always automatically
> imply that no idle balancing will happen, like
> on isolated CPUs?

IMO, nohz_full capable CPUs that are not isolated should automatically
become housekeepers, and nohz_full _active_ upon becoming isolated.
 When a used as a housekeeper, you still pay a price for having the
nohz_full capability available, but it doesn't have to be as high. 

In my kernels, I use cpusets to turn nohz on/off set wise, so CPUs can
be ticking, dyntick, nohz_full or housekeeper, RT load balancing and
cpupri on/off as well if you want to assume full responsibility.  It's
a tad (from box of xxl tads) ugly, but more flexible.

	-Mike