Messages in this thread |  | | | From | Nick Piggin <> | | Subject | Re: quicklists confuse meminfo | | Date | Tue, 11 Mar 2008 15:07:20 +1100 | |
On Tuesday 11 March 2008 02:54, Christoph Lameter wrote:
> On Sun, 9 Mar 2008, Andrew Morton wrote:
> > - They are pretty marginal from a performance POV (iirc)
>
> Pretty significant in our experience.
>
> > Making __GFP_ZERO smarter/more efficient would be a preferable way of
> > addressing any performance problems we have in there.
>
> Looking at the page allocator "fastpath": The basic reaons that this was
> such a good optimization was that the page allocator is expensive to call.
> The hotpath gets more and more clogged with logic. Fixing the page
> allocator to be more efficient may be the right approach here. That could
> also include having a list of zeroed pages.
This is insane. We add more and more of this NUMA and cpuset and anti-frag
and page zeroing logic to the page allocator, and then decide that we
don't actually need to obey any of those rules when we're running lat_proc.
You will never be able to make the page allocator faster than a single list
of pages. The reason is because we actually *want* some of these checks and
heuristics in the page allocator.
And I doubt a list of zeroed pages is the right approach. That's just
adding more complexity for lat_proc AFAIKS.
|  |