Re: [patch 4/4] sched: Distangle worker accounting from rq->lock

From: Ingo Molnar <mingo@elte.hu>
To: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>,
	LKML <linux-kernel@vger.kernel.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Jens Axboe <axboe@kernel.dk>,
	Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [patch 4/4] sched: Distangle worker accounting from rq->lock
Date: Thu, 23 Jun 2011 14:51:07 +0200	[thread overview]
Message-ID: <20110623125107.GB15430@elte.hu> (raw)
In-Reply-To: <BANLkTi=SxzG_fonb008KQzbZk16tuxm7NA@mail.gmail.com>

* Tejun Heo <tj@kernel.org> wrote:

> Hello, Ingo.
> 
> On Thu, Jun 23, 2011 at 12:44 PM, Ingo Molnar <mingo@elte.hu> wrote:
> >
> > * Tejun Heo <tj@kernel.org> wrote:
> >
> >> The patch description is simply untrue.  It does affect how wq
> >> behaves under heavy CPU load.  The effect might be perfectly okay
> >> but more likely it will result in subtle suboptimal behaviors under
> >> certain load situations which would be difficult to characterize
> >> and track down.  Again, the trade off (mostly killing of
> >> ttwu_local) could be okay but you can't get away with just claiming
> >> "there's no harm".
> >
> > Well, either it can be measured or not. If you can suggest a specific
> > testing method to Thomas, please do.
> 
> Crafting a test case where the suggested change results in worse 
> behavior isn't difficult (it ends up waking/creating workers which 
> it doesn't have to between ttwu and actual execution); however, as 
> with any micro benchmark, the problem is with assessing whether and 
> how much it would matter in actual workloads (whatever that means) 
> and workqueue is used throughout the kernel with widely varying 
> patterns making drawing conclusion a bit tricky. [...]

Well, please suggest a workload where it *matters* - as i suspect any 
workload tglx will come up with will have another 90% of workloads 
that you could suggest: so it's much better if you suggest a testing 
method.

When someone comes to me with a scheduler change i can give them the 
workloads that they should double check. See the changelog of this 
recent commit for example:

  c8b281161dfa: sched: Increase SCHED_LOAD_SCALE resolution

So please suggest a testing method.

> [...] Given that, changing the behavior for the worse just for this 
> cleanup doesn't sound like too sweet a deal.  Is there any other 
> reason to change the behavior (latency, whatever) other than the 
> ttuw_local ugliness?

Well, the ugliness is one aspect of it, but my main concern is simply 
testability: any claim of speedup or slowdown ought to be testable, 
right? I mean, we'd like to drive people towards coming with patches 
and number like Nikhil Rao did in c8b281161dfa, right?

Thanks,

	Ingo