Re: [RFC PATCH v2] nohz/sched: disable ilb on !mc_capable()

[Dominik Brodowski <linux@dominikbrodowski.net>]

On Tue, May 04, 2010 at 03:14:24PM +0200, Peter Zijlstra wrote:
> On Mon, 2010-04-26 at 22:31 +0200, Dominik Brodowski wrote:
> > From: Dominik Brodowski <linux@dominikbrodowski.net>
> > Date: Thu, 8 Apr 2010 21:51:18 +0200
> > Subject: [PATCH] nohz/sched: disable ilb on !mc_capable()
> > 
> > On my dual-core, !mc_capbale() CPU, the idle load balancer (ilb) is one
> > of the main reasons ticks are not stopped: Under moderate load (~98 % idle),
> > upt o half of the calls to tick_nohz_top_sched_tick() are aborted due
> > to calls to select_nohz_load_balancer(1).
> > 
> > I suspect this is caused by the following phenomenon:
> > 
> >     CPU0				CPU1
> >     <active>				<active>
> >     tick_nohz_stop_sched_tick(1)
> >     select_nohz_load_balancer(1)
> >      => CPU0 becomes ilb owner,		<CPU1 becomes idle a bit later>
> >         tick is not stopped,		tick_nohz_stop_sched_tick(1)
> >         CPU0 goes to sleep for		 => CPU1 isn't the ilb owner,
> >         exactly 1 tick.			    tick is stopped.
> >     <short sleep>			<long sleep>
> >     ---> scheduler_tick()
> >     tick_nohz_stop_sched_tick(0)
> >     tick_nohz_stop_sched_tick(1)
> >      => is ilb owner, all CPUs are
> >         idle, CPU0 may go to sleep.
> > 
> > If all CPU cores have hardly anything to do, letting the active CPU do
> > idle load balancing allows us to enter deep sleep states earlier, and for
> > longer periods of time. Furthermore, on !mc_capable() systems, it seems that
> > the ilb algorithm isn't needed at all. Let's show this for a 2-core system:
> > 
> > - if both cores are active, ilb is deactivated
> > - if no core is active, ilb is deactivated
> > - if only one core is active, it attempts to balance its load off to other
> >   CPUs on each tick anyway. ilb wouldn't act quicker.
> > 
> > This patch decreases the amount of wakeups on my completely idle notebook by
> > about two thirds.
> 
> Right, so I think the !mc_capable() check is buggy, at the very least on
> sparc64 which is 'creative' with its sched_domain maps.
> 
> I'm also not sure what a single socket AMD Magny-Cours will do.
> 
> On a single socket Nehalem we will have a non trivial sched_domain
> because we also have the threads included.
> 
> I think we can only do your optimization for machines that end up having
> a single sched_domain that covers the entire machine.

Is there an easy way to determine there's just a single sched_domain?

Best,
	Dominik