Re: [PATCH] sched: wakeup buddy

[Michael Wang <wangyun@linux.vnet.ibm.com>]

On 03/14/2013 06:58 PM, Peter Zijlstra wrote:
> On Wed, 2013-03-13 at 11:07 +0800, Michael Wang wrote:
> 
>> However, we already figure out the logical that wakeup related task
>> could benefit from closely running, this could promise us somewhat
>> reliable benefit.
> 
> I'm not convinced that the 2 task wakeup scenario is the only sane
> scenario. Imagine a ring of 3 tasks that wakes each other, if the
> machine is loaded enough, those 3 tasks might fit a single cpu just
> fine -- esp. if one of those is always idle.
> 
> But your strict 1:1 wakeup relation thing will completely fail to
> detect this.

Hmm...yeah, I see your point here, some optimize timing will always
be missed whatever how we twiddle the knob besides 0.

You are right, that's a problem, although currently there are no
workload to prove it, but we have the theory...

> 
>> IMHO, that sounds a little easier for users than to make the decision on
>> tell me how long to pull tasks together, they may be confused...
> 
> Users shouldn't ever need/want/etc.. rely on this. They should just run
> their programs and be (reasonably) happy.
> 
>> In summary, I think we have two point here need to be considered:
>>
>> 1. what about the missed optimize timing, that may benefit
>>    some workload (although we haven't found such workload yet).
> 
> Missed optimize; as in not calling wake_affine() due to the throttle?
> If we're calling it at such a high frequency it is very likely the next
> call isn't very far away.
> 
>> 2. how many benefit could wake_affine() stuff bring to us,
>>    if limit rate benefit us, why don't we remove it?
> 
> It could bring the same benefit but at lower overhead, what's the point
> of computing the same value over and over again? Also, the rate limit
> thing naturally works for the soft/hard-irq case.

Just try to confirm my understanding, so we are going to do something
like:

	if (now - wakee->last > time_limit) && wakeup_affine()
		wakee->last = now
		select_idle_sibling(curr_cpu)
	else
		select_idle_sibling(prev_cpu)

And time_limit is some static value respect to the rate of load balance,
is that correct?

Currently I haven't found regression by reduce the rate, but if we found
such benchmark, we may still need a way (knob or CONFIG) to disable this
limitation.

> 
> Now, there's another detail I thought up, one could only limit the
> wake_affine() calls once it starts returning false.

Hmm..if wake_affine() keep succeed, then there will be no difference?

I do believe pgbench match the case, since wake_affine() stuff make it
suffered...and the more it suffered, means the more often wake_affine()
succeed and pull none related tasks together.

I really can't see how could it do help... did I miss something?

Regards,
Michael Wang

>