Re: kernbench: 2.5.59 vs 2.5.59-mm6 vs 2.5.59-mjb2

[William Lee Irwin III]

On Wed, Jan 29, 2003 at 02:13:31PM -0800, Martin J. Bligh wrote:
> not sure what's causing the pfn_to_nid and pgd_alloc improvements?)

pgd's and pmd's are in the slab, so the zeroing hit is only taken once
per pgd and pmd (best case). It's my patch that does this.

pfn_to_nid() is surely a random cache and/or ITLB gain from link
ordering, alignment in the final link, or some such nonsense.
pfn_to_nid() is a single shift, quite excessive for a function call.
It'd be best to inline this and get the random ITLB etc. misses out
of the profiles esp. as they aren't reproducible.

-- wli