Re: [PATCH V3]hrtimer: Fix a performance regression by disablereprogramming in remove_hrtimer

[Mike Galbraith]

On Thu, 2013-08-08 at 23:02 +0800, ethan.zhao wrote: 
> å 2013-8-6ïäå3:29ïMike Galbraith <bitbucket@xxxxxxxxx> åéï
> 
> > +int sched_needs_cpu(int cpu)
> > +{
> > +	return  cpu_rq(cpu)->avg_idle < sysctl_sched_migration_cost;
> > +}
> > +
> > #else /* CONFIG_NO_HZ_COMMON */
> > 
> > static inline bool got_nohz_idle_kick(void)
> > --- a/kernel/time/tick-sched.c
> > +++ b/kernel/time/tick-sched.c
> > @@ -548,7 +548,7 @@ static ktime_t tick_nohz_stop_sched_tick
> > 		time_delta = timekeeping_max_deferment();
> > 	} while (read_seqretry(&jiffies_lock, seq));
> > 
> > -	if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
> > +	if (sched_needs_cpu(cpu) || rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
> > 	    arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
> > 		next_jiffies = last_jiffies + 1;
> > 		delta_jiffies = 1;
> 
> If the performace regression was caused by too much expensive clock device reprogramming and too frequent entering /exiting of C-statesâ   this patch should work.
> except the following result is almost always false under 3.11-rc3 code.
> 
> > return  cpu_rq(cpu)->avg_idle < sysctl_sched_migration_cost;

On my E5620 box, avg_idle works fine.  Patchlet doesn't save as much as
it used to, thanks to Peter's patch now killing the worst of the pain,
but it does still does save cycles.  I have too much regression left to
say exactly what it can now save max, doesn't matter much either.

The pertinent numbers:

v3.11-rc4-27-ge4ef108    496.7 KHz      .858     1.030     throttle+peterz
v3.11-rc4-27-ge4ef108    440.7 KHz      .761     1.296     nothrottle+peterz

-Mike

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/