On Tue, 20 Aug 2019, Shakeel Butt wrote:
> On Tue, Aug 20, 2019 at 3:45 AM Michal Hocko <mhocko@kernel.org> wrote:
> > On Tue 20-08-19 17:48:23, Alex Shi wrote:
> > > This patchset move lru_lock into lruvec, give a lru_lock for each of
> > > lruvec, thus bring a lru_lock for each of memcg.
> > >
> > > Per memcg lru_lock would ease the lru_lock contention a lot in
> > > this patch series.
> > >
> > > In some data center, containers are used widely to deploy different kind
> > > of services, then multiple memcgs share per node pgdat->lru_lock which
> > > cause heavy lock contentions when doing lru operation.
> >
> > Having some real world workloads numbers would be more than useful
> > for a non trivial change like this. I believe googlers have tried
> > something like this in the past but then didn't have really a good
> > example of workloads that benefit. I might misremember though. Cc Hugh.
> >
> 
> We, at Google, have been using per-memcg lru locks for more than 7
> years. Per-memcg lru locks are really beneficial for providing
> performance isolation if there are multiple distinct jobs/memcgs
> running on large machines. We are planning to upstream our internal
> implementation. I will let Hugh comment on that.

Thanks for the Cc Michal.  As Shakeel says, Google prodkernel has been
using our per-memcg lru locks for 7 years or so.  Yes, we did not come
up with supporting performance data at the time of posting, nor since:
I see Alex has done much better on that (though I haven't even glanced
to see if +s are persuasive).

https://lkml.org/lkml/2012/2/20/434
was how ours was back then; some parts of that went in, then attached
lrulock417.tar is how it was the last time I rebased, to v4.17.

I'll set aside what I'm doing, and switch to rebasing ours to v5.3-rc
and/or mmotm.  Then compare with what Alex has, to see if there's any
good reason to prefer one to the other: if no good reason to prefer ours,
I doubt we shall bother to repost, but just use it as basis for helping
to review or improve Alex's.

Hugh