Re: [PATCH 1/4] mm: rearrange node_stat_item to put memcg stats at start

[Chris Li <chrisl@kernel.org>]

On Tue, Apr 23, 2024 at 11:30 AM Johannes Weiner <hannes@cmpxchg.org> wrote:
>
> On Tue, Apr 23, 2024 at 10:44:07AM -0700, Shakeel Butt wrote:
> > On Tue, Apr 23, 2024 at 09:58:44AM -0400, Johannes Weiner wrote:
> > > On Mon, Apr 22, 2024 at 10:18:23PM -0700, Shakeel Butt wrote:
> > > > At the moment the memcg stats are sized based on the size of enum
> > > > node_stat_item but not all fields in node_stat_item corresponds to memcg
> > > > stats. So, rearrage the contents of node_stat_item such that all the
> > > > memcg specific stats are at the top and then the later patches will make
> > > > sure that the memcg code will not waste space for non-memcg stats.
> > > >
> > > > Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
> > >
> > > This series is a great idea and the savings speak for themselves.
> > >
> > > But rearranging and splitting vmstats along the memcg-nomemcg line
> > > seems like an undue burden on the non-memcg codebase and interface.
> > >
> > > - It messes with user-visible /proc/vmstat ordering, and sets things
> > >   up to do so on an ongoing basis as stats are added to memcg.
> > >
> > > - It also separates related stats (like the workingset ones) in
> > >   /proc/vmstat when memcg only accounts a subset.
> > >
> > > Would it make more sense to have a translation table inside memcg?
> > > Like we have with memcg1_events.
> >
> > Thanks for taking a look. I will look into the translation table
> > approach. The reason I went with this approach was that I am in parallel
> > looking into rearranging fields of important MM structs and also enums
> > to improve cache locality. For example, the field NR_SWAPCACHE is always
> > accessed together with NR_FILE_PAGES, so it makes sense to have them on
> > same cacheline. So, is the rearrangement of vmstats a big NO or a little
> > bit here and there is fine unlike what I did with this patch?
>
> I'm curious what other folks think.
>
I slightly prefer not to change user visible ordering for no good reason.
It is not said the order is carved to stone. It depends on the ROI.

> The cache optimization is a stronger argument, IMO, because it
> directly benefits the users of /proc/vmstat. And it would be fairly
> self contained inside the node_stat_item enum - "ordered for cache".

Not sure how much of the cache optimization is measurable here.
I suspect it is going to be hard to measure a meaningful difference
just from the cache line order alone.

> I was more hesitant about imposing a memcg requirement on the generic
> vmstat ordering.

That is a valid reason.

Chris