Date:	Sat, 05 Apr 2014 04:01:30 +0200
From:	"Rafael J. Wysocki" <rjw@...ysocki.net>
To:	Nicolas Pitre <nicolas.pitre@...aro.org>
Cc:	Daniel Lezcano <daniel.lezcano@...aro.org>,
	linux-kernel@...r.kernel.org, mingo@...e.hu, peterz@...radead.org,
	linux-pm@...r.kernel.org, alex.shi@...aro.org,
	vincent.guittot@...aro.org, morten.rasmussen@....com
Subject: Re: [RFC PATCHC 3/3] sched/fair: use the idle state info to choose the idlest cpu

On Friday, April 04, 2014 12:56:52 PM Nicolas Pitre wrote:
> On Fri, 4 Apr 2014, Rafael J. Wysocki wrote:
> 
> > On Tuesday, April 01, 2014 11:05:49 PM Nicolas Pitre wrote:
> > > On Fri, 28 Mar 2014, Daniel Lezcano wrote:
> > > 
> > > > As we know in which idle state the cpu is, we can investigate the following:
> > > > 
> > > > 1. when did the cpu entered the idle state ? the longer the cpu is idle, the
> > > > deeper it is idle
> > > > 2. what exit latency is ? the greater the exit latency is, the deeper it is
> > > > 
> > > > With both information, when all cpus are idle, we can choose the idlest cpu.
> > > > 
> > > > When one cpu is not idle, the old check against weighted load applies.
> > > > 
> > > > Signed-off-by: Daniel Lezcano <daniel.lezcano@...aro.org>
> > > 
> > > There seems to be some problems with the implementation.
> > > 
> > > > @@ -4336,20 +4337,53 @@ static int
> > > >  find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
> > > >  {
> > > >  	unsigned long load, min_load = ULONG_MAX;
> > > > -	int idlest = -1;
> > > > +	unsigned int min_exit_latency = UINT_MAX;
> > > > +	u64 idle_stamp, min_idle_stamp = ULONG_MAX;
> > > 
> > > I don't think you really meant to assign an u64 variable with ULONG_MAX.
> > > You probably want ULLONG_MAX here.  And probably not in fact (more 
> > > later).
> > > 
> > > > +
> > > > +	struct rq *rq;
> > > > +	struct cpuidle_power *power;
> > > > +
> > > > +	int cpu_idle = -1;
> > > > +	int cpu_busy = -1;
> > > >  	int i;
> > > >  
> > > >  	/* Traverse only the allowed CPUs */
> > > >  	for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
> > > > -		load = weighted_cpuload(i);
> > > >  
> > > > -		if (load < min_load || (load == min_load && i == this_cpu)) {
> > > > -			min_load = load;
> > > > -			idlest = i;
> > > > +		if (idle_cpu(i)) {
> > > > +
> > > > +			rq = cpu_rq(i);
> > > > +			power = rq->power;
> > > > +			idle_stamp = rq->idle_stamp;
> > > > +
> > > > +			/* The cpu is idle since a shorter time */
> > > > +			if (idle_stamp < min_idle_stamp) {
> > > > +				min_idle_stamp = idle_stamp;
> > > > +				cpu_idle = i;
> > > > +				continue;
> > > 
> > > Don't you want the highest time stamp in order to select the most 
> > > recently idled CPU?  Favoring the CPU which has been idle the longest 
> > > makes little sense.
> > 
> > It may make sense if the hardware can auto-promote CPUs to deeper C-states.
> 
> If so the promotion will happen over time, no?  What I'm saying here is 
> that those CPUs which have been idle longer should not be favored when 
> it is time to select a CPU for a task to run. More recently idled CPUs 
> are more likely to be in a shallower C-state.
> 
> > Something like that happens with package C-states that are only entered when
> > all cores have entered a particular core C-state already.  In that case the
> > probability of the core being in a deeper state grows with time.
> 
> Exactly what I'm saying.

Right, I got that the other way around by mistake.

> Also here it is worth remembering that the scheduling domains should 
> represent those packages that share common C-states at a higher level.  
> The scheduler can then be told not to balance across domains if it 
> doesn't need to in order to favor the conditions for those package 
> C-states to be used.  That's what the task packing patch series is 
> about, independently of this one.
> 
> > That said I would just drop this heuristics for the time being.  If auto-promotion
> > is disregarded, it doesn't really matter how much time the given CPU has been idle
> > except for one case: When the target residency of its idle state hasn't been
> > reached yet, waking up the CPU may be a mistake (depending on how deep the state
> > actually is, but for the majority of drivers in the tree we don't have any measure
> > of that).
> 
> There is one reason for considering the time a CPU has been idle, 
> assuming equivalent C-state, and that is cache snooping.  The longer a 
> CPU is idle, the more likely its cache content will have been claimed 
> and migrated by other CPUs.  Of course that doesn't make much difference 
> for deeper C-states where the cache isn't preserved, but it is probably 
> simpler and cheaper to apply this heuristic in all cases.

Yes, that sounds like it might be a reason, but I'd like to see numbers
confirming that to be honest.

> > > > +			}
> > > > +
> > > > +			/* The cpu is idle but the exit_latency is shorter */
> > > > +			if (power && power->exit_latency < min_exit_latency) {
> > > > +				min_exit_latency = power->exit_latency;
> > > > +				cpu_idle = i;
> > > > +				continue;
> > > > +			}
> > > 
> > > I think this is wrong.  This gives priority to CPUs which have been idle 
> > > for a (longer... although this should have been) shorter period of time 
> > > over those with a shallower idle state.  I think this should rather be:
> > > 
> > > 	if (power && power->exit_latency < min_exit_latency) {
> > > 		min_exit_latency = power->exit_latency;
> > > 		latest_idle_stamp = idle_stamp;
> > > 	       	cpu_idle = i;
> > > 	} else if ((!power || power->exit_latency == min_exit_latency) &&
> > > 		   idle_stamp > latest_idle_stamp) {
> > > 		latest_idle_stamp = idle_stamp;
> > > 		cpu_idle = i;
> > > 	}
> > > 
> > > So the CPU with the shallowest idle state is selected in priority, and 
> > > if many CPUs are in the same state then the time stamp is used to 
> > > select the most recent one.
> > 
> > Again, if auto-promotion is disregarded, it doesn't really matter which of them
> > is woken up.
> 
> If it doesn't matter then it doesn't hurt.  But in some cases it 
> matters.
> 
> > > Whenever a shallower idle state is found then the latest_idle_stamp is reset for 
> > > that state even if it is further in the past.
> > > 
> > > > +		} else {
> > > > +
> > > > +			load = weighted_cpuload(i);
> > > > +
> > > > +			if (load < min_load ||
> > > > +			    (load == min_load && i == this_cpu)) {
> > > > +				min_load = load;
> > > > +				cpu_busy = i;
> > > > +				continue;
> > > > +			}
> > > >  		}
> > > 
> > > I think this is wrong to do an if-else based on idle_cpu() here.  What 
> > > if a CPU is heavily loaded, but for some reason it happens to be idle at 
> > > this very moment?  With your patch it could be selected as an idle CPU 
> > > while it would be discarded as being too busy otherwise.
> > 
> > But see below ->
> > 
> > > It is important to determine both cpu_busy and cpu_idle for all CPUs.
> > > 
> > > And cpu_busy is a bad name for this.  Something like least_loaded would 
> > > be more self explanatory.  Same thing for cpu_idle which could be 
> > > clearer if named shalloest_idle.
> > 
> > shallowest_idle?
> 
> Something that means the CPU with the shallowest C-state.  Using 
> "cpu_idle" for this variable doesn't cut it.

Yes, that was about the typo above only. :-)

> > > > -	return idlest;
> > > > +	/* Busy cpus are considered less idle than idle cpus ;) */
> > > > +	return cpu_busy != -1 ? cpu_busy : cpu_idle;
> > > 
> > > And finally it is a policy decision whether or not we want to return 
> > > least_loaded over shallowest_idle e.g do we pack tasks on non idle CPUs 
> > > first or not.  That in itself needs more investigation.  To keep the 
> > > existing policy unchanged for now the above condition should have its 
> > > variables swapped.
> > 
> > Which means that once we've find the first idle CPU, it is not useful to
> > continue computing least_loaded, because we will return the idle one anyway,
> > right?
> 
> Good point.  Currently, that should be the case.
> 
> Eventually we'll want to put new tasks on lightly loaded CPUs instead of 
> waking up a fully idle CPU in order to favor deeper C-states. But that 
> requires a patch series of its own just to determine how loaded a CPU is 
> and how much work it can still accommodate before being oversubscribed, 
> etc.

Wouldn't we need power consumption numbers for that realistically?

Rafael

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

Powered by blists - more mailing lists