Re: AIM7 40% regression with 2.6.26-rc1

[Ingo Molnar <mingo@elte.hu>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> > I think it is far more likely that it's due to the different 
> > scheduling and wakeup behavior of the new kernel/semaphore.c code. 
> > So the fix would be to restore the old scheduling behavior - that's 
> > what Yanmin's manual revert did and that's what got him back the 
> > previous AIM7 performance.
> 
> Yes, Yanmin's manual revert got rid of the new semaphores entirely. 
> Which was what, 7500 lines of code removed that got reverted.

i wouldnt advocate a 7500 revert instead of a 160 lines change.

my suggestion was that the scheduling behavior of the new 
kernel/semaphore.c code is causing the problem - i.e. making it match 
the old semaphore code's behavior would give us back performance.

> And the *WHOLE* and *ONLY* excuse for dropping the spinlock 
> lock_kernel was this (and I quote your message):
> 
>     remove the !PREEMPT_BKL code.
>     
>     this removes 160 lines of legacy code.
> 
> in other words, your only stated valid reason for getting rid of the 
> spinlock was 160 lines, and the comment didn't even match what it did 
> (it removed the spinlocks entirely, not just the preemptible version).

it was removed by me in the course of this discussion:

   http://lkml.org/lkml/2008/1/2/58

the whole discussion started IIRC because !CONFIG_PREEMPT_BKL [the 
spinlock version] was broken for a longer period of time (it crashed 
trivially), because nobody apparently used it. People (Nick) asked why 
it was still there and i agreed and removed it. CONFIG_PREEMPT_BKL=y was 
the default, that was what all distros used. I.e. the spinlock code was 
in essence dead code at that point in time.

the spinlock code might in fact perform _better_, but nobody came up 
with such a workload before.

> In contrast, the revert adds 7500 lines. If you go by the only 
> documented reason for the crap that is the current BKL, then I know 
> which one I'll take. I'll take the spinlock back, and I'd rather put 
> preemption back than ever take those semaphores.
> 
> And even that's ignoring another issue: did anybody ever even do that 
> AIM7 benchmark comparing spinlocks to the semaphore-BKL? It's quite 
> possible that the semaphores (even the well-behaved ones) behaved 
> worse than the spinlocks.

that's a good question...

	Ingo