Re: [rfc] increase struct page size?!

[Nick Piggin <npiggin@suse.de>]

On Mon, May 21, 2007 at 01:08:13AM -0700, William Lee Irwin III wrote:
> On Sun, May 20, 2007 at 01:46:47AM -0700, William Lee Irwin III wrote:
> >> The lack of consideration of the average case. I'll see what I can smoke
> >> out there.
> 
> On Sun, May 20, 2007 at 11:25:52AM +0200, Nick Piggin wrote:
> > I _am_ considering the average case, and I consider the aligned structure
> > is likely to win on average :) I just don't have numbers for it yet.
> 
> Choosing k distinct integers (mem_map array indices) from the interval
> [0,n-1] results in k(n-k+1)/n non-adjacent intervals of contiguous
> array indices on average. The average interval length is
> (n+1)/(n-k+1) - 1/C(n,k). Alignment considerations make going much
> further somewhat hairy, but it should be clear that contiguity arising
> from random choice is non-negligible.

That doesn't say anything about temporal locality, though.

 
> In any event, I don't have all that much of an objection to what's
> actually proposed, just this particular cache footprint argument.
> One can motivate increases in sizeof(struct page), but not this way.

Realise that you have to have a run of I think at least 7 or 8 contiguous
pages and temporally close references in order to save a single cacheline.

Then also that if the page being touched is not partially in cache from
an earlier access, then it is statistically going to cost more lines to
touch it (up to 75% if you touch the first and the last field, obviously 0%
if you only touch a single field, but that's unlikely given that you
usually take a reference then do at least something else like check flags).

I think the problem with the cache footprint argument is just whether
it makes any significant difference to performance. But..


> Now that I've been informed of the ->_count and ->_mapcount issues,
> I'd say that they're grave and should be corrected even at the cost
> of sizeof(struct page).

... yeah, something like that would bypass 
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/