Re: [PATCH 2/2] slub: fix Objects count

[Hugh Dickins <hugh@veritas.com>]

On Sat, 3 Nov 2007, Christoph Lameter wrote:
> On Sat, 3 Nov 2007, Hugh Dickins wrote:
> 
> > I was afraid you might say something like that.
> > Perhaps it'll be a patch I need to use in my own builds.
> > Though I'd have thought others would want that accuracy too.
> > Didn't SLAB give it?  (The "r*gr*ss**n" word!)
> 
> Slab also only counts objects that are not in the queues. See free_block() 
> f.e.

I'll take your word for it, and apologize for my slur on slub!
(Slub has a great deal to admire in it, I should say.)

> 
> We could improve the situation by flushing all cpu slabs before counts are 
> determined.
> 
> Which can be done manually. Run
> 
> 	slabinfo -s
> 
> and then look at the numbers.

Mmm, I'd been doing slabinfo -v sometimes.  These are fine in some
situations, but it's always better when the observer can avoid
interfering with the observed.  Impossible, we know, but...

Also, many caches too quickly re-equip themselves
with cpu slabs which again obscure the numbers.

> > > Adds to much overhead to the fast paths
> > 
> > You've come to that conclusion very quickly!
> 
> I have just spend a few weeks optimizing the fast and slow paths and there 
> is some additional overhead that I am still trying to eliminate.
> 
> > Any numbers to back it up?
> 
> The performance in the fast paths depends on updating only a single word
> for an allocation. Adding another counter makes that impossible.

Gosh, that's a tighter corner than any I've been in.

> 
> See the recent post on SLUB regression on SMP.

I'll have to read up on that, thanks for the pointer.

Hugh