Re: [RFC] Convert pgtable cache to slab

[Robin Holt <holt@sgi.com>]

On Thu, Oct 14, 2004 at 12:44:59AM -0700, David Mosberger wrote:
> >>>>> On Wed, 13 Oct 2004 11:59:56 -0700, William Lee Irwin III <wli@holomorphy.com> said:
> 
>   William> On Wed, Oct 13, 2004 at 02:47:43PM -0400, Martin
>   William> K. Petersen wrote:
>   >> This patch makes the page table cache on IA-64 use the slab
>   >> instead of standard page allocations.  It's based upon Bill
>   >> Irwin's code for ppc64.  With this patch I got a significant
>   >> improvement in page fault time.  Standard 2.4.x was about 700ns
>   >> on average.  Generic 2.6.9rc4 is 3-4 usec, whereas the slabified
>   >> pgtcache drops us back down to 600-700 ns.  Tested on zx1 and
>   >> sn2.  The biggest caveat is that I've had to postpone setting up
>   >> the gate page until the pgt slab has been initialized.  That
>   >> wasn't an issue with the existing page-based allocation scheme.
>   >> David - how do you prefer I handle this?
> 
>   William> Nice! I got reports that it would not be beneficial when I
>   William> thought about going over this earlier. I suppose it's a
>   William> small vindication of my methods to see the original
>   William> objection contradicted here. =)
> 
> What was the original comment relative to?  2.4 or 2.6?  In 2.6 we
> lost the PTE quick-list, which I think is where the big overhead came
> from.  Right, Martin?

And the pmd quicklist.  We add entries to the pmd quicklist, but never
take them back off.  This has a two fold effect in that it causes the
pgd quicklist to be drained unnecessarily when the pmd quicklist has
many unusable entries on it.

Essentially, we were allocating and clearing pages for both pmd and pte
entries every time.

Robin
> 
> 	--david
> -
> To unsubscribe from this list: send the line "unsubscribe linux-ia64" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html