Re: AIM7 40% regression with 2.6.26-rc1

[Linus Torvalds]

On Thu, 8 May 2008, Zhang, Yanmin wrote:
> 
> On Tue, 2008-05-06 at 10:23 -0600, Matthew Wilcox wrote:
> > On Tue, May 06, 2008 at 06:09:34AM -0600, Matthew Wilcox wrote:
> > > So the only likely things I can see are:
> > > 
> > >  - file locks
> > >  - fasync
> > 
> > I've wanted to fix file locks for a while.  Here's a first attempt.
> > It was done quickly, so I concede that it may well have bugs in it.
> > I found (and fixed) one with LTP.
> > 
> > It takes *no account* of nfsd, nor remote filesystems.  We need to have
> > a serious discussion about their requirements.
>
> I tested it on 8-core stoakley. aim7 result becomes 23% worse than the one of
> pure 2.6.26-rc1.

Ouch. That's really odd. The BKL->spinlock conversion looks really 
obvious, so it shouldn't be that noticeably slower.

The *one* difference is that the BKL has the whole "you can take it 
recursively and you can sleep without dropping it because the scheduler 
will drop it for you" thing. The spinlock conversion changed all of that 
into explicit "drop and retake" locks, and maybe that causes some issues. 

But 23% worse? That sounds really odd/extreme.

Can you do a oprofile run or something?

			Linus
--
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/