Re: [patch 1/3] mm: embed the memcg pointer directly into struct page

[David Miller <davem-fT/PcQaiUtIeIZ0/mPfg9Q@public.gmane.org>]

From: "Kirill A. Shutemov" <kirill-oKw7cIdHH8eLwutG50LtGA@public.gmane.org>
Date: Mon, 3 Nov 2014 23:52:06 +0200

> On Mon, Nov 03, 2014 at 04:36:28PM -0500, Johannes Weiner wrote:
>> On Mon, Nov 03, 2014 at 11:06:07PM +0200, Kirill A. Shutemov wrote:
>> > On Sat, Nov 01, 2014 at 11:15:54PM -0400, Johannes Weiner wrote:
>> > > Memory cgroups used to have 5 per-page pointers.  To allow users to
>> > > disable that amount of overhead during runtime, those pointers were
>> > > allocated in a separate array, with a translation layer between them
>> > > and struct page.
>> > > 
>> > > There is now only one page pointer remaining: the memcg pointer, that
>> > > indicates which cgroup the page is associated with when charged.  The
>> > > complexity of runtime allocation and the runtime translation overhead
>> > > is no longer justified to save that *potential* 0.19% of memory.
>> > 
>> > How much do you win by the change?
>> 
>> Heh, that would have followed right after where you cut the quote:
>> with CONFIG_SLUB, that pointer actually sits in already existing
>> struct page padding, which means that I'm saving one pointer per page
>> (8 bytes per 4096 byte page, 0.19% of memory), plus the pointer and
>> padding in each memory section.  I also save the (minor) translation
>> overhead going from page to page_cgroup and the maintenance burden
>> that stems from having these auxiliary arrays (see deleted code).
> 
> I read the description. I want to know if runtime win (any benchmark data?)
> from moving mem_cgroup back to the struct page is measurable.
> 
> If the win is not significant, I would prefer to not occupy the padding:
> I'm sure we will be able to find a better use for the space in struct page
> in the future.

I think the simplification benefits completely trump any performan
metric.