Re: [RFC PATCH] sched: smart wake-affine

[Michael Wang]

On 06/03/2013 11:53 AM, Mike Galbraith wrote:
> On Mon, 2013-06-03 at 11:26 +0800, Michael Wang wrote: 
>> On 06/03/2013 11:09 AM, Mike Galbraith wrote:
>>> On Mon, 2013-06-03 at 10:28 +0800, Michael Wang wrote: 
>>>> On 05/28/2013 01:05 PM, Michael Wang wrote:
>>>>> wake-affine stuff is always trying to pull wakee close to waker, by theory,
>>>>> this will bring benefit if waker's cpu cached hot data for wakee, or the
>>>>> extreme ping-pong case.
>>>>>
>>>>> And testing show it could benefit hackbench 15% at most.
>>>>>
>>>>> However, the whole stuff is somewhat blindly and time-consuming, some
>>>>> workload therefore suffer.
>>>>>
>>>>> And testing show it could damage pgbench 50% at most.
>>>>>
>>>>> Thus, wake-affine stuff should be smarter, and realise when to stop
>>>>> it's thankless effort.
>>>>
>>>> Is there any comments?
>>>
>>> (I haven't had time to test-drive yet, -rt munches time like popcorn)
>>
>> I see ;-)
>>
>> During my testing, this one works well on the box, solved the issues of
>> pgbench and won't harm hackbench any, I think we have caught some good
>> point here :)
> 
> Some wider spectrum testing needs doing though.

That's right, the benchmark I currently have is hackbench, pgbench,
ebizzy, aim7, tbench, dbench, kbench, is there any other good candidate
we should add to the test?

Hackbench is a good
> sign, but localhost and db type stuff that really suffer from misses
> would be good to test.  Java crud tends to be sensitive too.  I used to
> watch vmark (crap) as an indicator,

I can't get it from google...do you mean vmmark?

if you see unhappiness there, you'll
> very likely see it in other loads as well, it is very fond of cache
> affine wakeups, but loathes preemption (super heavy loads usually do).

I agree that this idea, in other work, 'stop wake-affine when current is
busy with wakeup' may miss the chance to bring benefit, although I could
not find such workload, but I can't do promise...

However, IMHO, wake-affine is actually an additional accelerator, if we
want to adopt it without any knob, what we have to make sure is, it
won't damage any workload, rather than it catch all the chances to bring
benefit.

Here my point is, we stop the additional accelerator if it seems to
bring damage, and we currently found one such case ;-)

So the worst result of this idea is missing the timing to accelerate,
however, since such acceleration will sacrifice others and we don't want
knob, just filter it out.

Regards,
Michael Wang

> 
> -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/
> 

--
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/