Re: [PATCH] mm, page_alloc: skip zone who has no managed_pages in calculate_totalreserve_pages()

[Michal Hocko <mhocko@suse.com>]

On Mon 12-11-18 14:26:41, Wei Yang wrote:
> On Mon, Nov 12, 2018 at 09:09:26AM +0100, Michal Hocko wrote:
> >On Mon 12-11-18 15:14:04, Wei Yang wrote:
> >> Zone with no managed_pages doesn't contribute totalreserv_pages. And the
> >> more nodes we have, the more empty zones there are.
> >> 
> >> This patch skip the zones to save some cycles.
> >
> >What is the motivation for the patch? Does it really cause any
> >measurable difference in performance?
> >
> 
> The motivation here is to reduce some unnecessary work.

I have guessed so even though the changelog was quite modest on the
motivation.

> Based on my understanding, almost every node has empty zones, since
> zones within a node are ordered in monotonic increasing memory address.

Yes, this is likely the case. Btw. a check for populated_zone or
for_each_populated_zone would suite much better.

> The worst case is all zones has managed_pages. For example, there is
> only one node, or configured to have only ZONE_NORMAL and
> ZONE_MOVABLE. Otherwise, the more node/zone we have, the more empty
> zones there are.
> 
> I didn't have detail tests on this patch, since I don't have machine
> with large numa nodes. While compared with the following ten lines of
> code, this check to skip them is worthwhile to me.

Well, the main question is whether the optimization is really worth it.
There is not much work done for each zone.

I haven't looked closer whether the patch is actually correct, it seems
to be though, but optimizations without measurable effect tend to be not
that attractive.

-- 
Michal Hocko
SUSE Labs