Re: Progress on scaling of FreeBSD on 8 CPU systems

On Sunday 25 February 2007 05:56, Robert Watson wrote:
> On Sun, 25 Feb 2007, Ivan Voras wrote:
> 
> > Kris Kennaway wrote:
> >
> >> Hopefully within a week or two.  It might not be that exact patch, I
> >> think John wants to try and do it a bit differently instead of
> >> introducing a new locking primitive just for this.
> >
> > Well why not? :) I am not an expert, but reading jeffr's posts it looks 
like 
> > the idea of sleepable mutexes was taken from Solaris, where it's also not 
> > exactly documented. If moving away from sleepable mutexes introduces more 
> > than a small single digit percentage drop in performance (1% on 
> > multi-gigahertz machines is a lot), why not keep it? If it's dangerous to 
> > use, that should be documented in the man page with big bold letters but 
if 
> > it helps, keep it.
> >
> > (Of course I might be completely off the track and sleepable mutexes might 
> > be inconsequential for performance here :) )
> 
> Well, there are two ways you can ask the question about locks here:
> 
> (1) Why don't we allow sleeping with mutexes?
> 
> (2) Why don't the sleepable locking primitives perform better?
> 
> There are now patches that address this from both sides, optimizing sx lock 
> performance and allowing mutexes to sleep.  There are serious deadlock 
issues 
> that can arise with sleepable mutexes; I believe Jeff's patch includes the 
> necessary bits to teach WITNESS how to detect some misuse at run-time. Right 
> now, with the exception of the fast interrupt context, mutexes are 
universally 
> acquirable in any context subject to lock order.  If we have sleepable 
> mutexes, this will no longer be true, which is a significant change in the 
> constraints on use. Attilio has a heavily optimized sxlock implementation as 
> well, although I'm not sure the two have been benchmarked side-by-side, but 
> that would be an obvious next thing to try.

My intention is to both get the sx locks changes into the tree so that sx 
locks use atomic ops, and to add a new sleep lock primitive that is like a 
mutex but uses sleepq rather than turnstiles.  The latter is a lot like 
Jeff's top-half mutex just with some slight tweaks, and as a separate lock 
primitive.

-- 
John Baldwin