Re: [RFC PATCH] memcg: use root_mem_cgroup when css is inherited

[Zhaoyang Huang]

On Tue, Aug 23, 2022 at 1:21 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
>
> On Tue 23-08-22 10:31:57, Zhaoyang Huang wrote:
> > On Mon, Aug 22, 2022 at 7:31 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >
> > > On Fri 19-08-22 07:10:26, Tejun Heo wrote:
> > > > On Fri, Aug 19, 2022 at 10:08:59AM -0700, Shakeel Butt wrote:
> > > > > On Fri, Aug 19, 2022 at 9:29 AM Tejun Heo <tj@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Fri, Aug 19, 2022 at 07:29:22PM +0800, zhaoyang.huang wrote:
> > > > > > > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > > > > >
> > > > > > > It is observed in android system where per-app cgroup is demanded by freezer
> > > > > > > subsys and part of groups require memory control. The hierarchy could be simplized
> > > > > > > as bellowing where memory charged on group B abserved while we only want have
> > > > > > > group E's memory be controlled and B's descendants compete freely for memory.
> > > > > > > This should be the consequences of unified hierarchy.
> > > > > > > Under this scenario, less efficient memory reclaim is observed when comparing
> > > > > > > with no memory control. It is believed that multi LRU scanning introduces some
> > > > > > > of the overhead. Furthermore, page thrashing is also heavier than global LRU
> > > > > > > which could be the consequences of partial failure of WORKINGSET mechanism as
> > > > > > > LRU is too short to protect the active pages.
> > > > > > >
> > > > > > > A(subtree_control = memory) - B(subtree_control = NULL) - C()
> > > > > > >                                                       \ D()
> > > > > > >                           - E(subtree_control = memory) - F()
> > > > > > >                                                         \ G()
> > > > > > >
> > > > > > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > > > >
> > > > > > Just in case it wasn't clear.
> > > > > >
> > > > > > Nacked-by: Tejun Heo <tj@xxxxxxxxxx>
> > > > > >
> > > > > > Thanks.
> > > > > >
> > > > >
> > > > > Was there a previous discussion on this? The commit message is unreadable.
> > > >
> > > > http://lkml.kernel.org/r/1660298966-11493-1-git-send-email-zhaoyang.huang@xxxxxxxxxx
> > >
> > > Even that discussion doesn't really explain the real underlying problem.
> > > There are statements about inefficiency and trashing without any further
> > > details or clarifications.
> > I would like to quote the comments from google side for more details
> > which can also be observed from different vendors.
> > "Also be advised that when you enable memcg v2 you will be using
> > per-app memcg configuration which implies noticeable overhead because
> > every app will have its own group. For example pagefault path will
> > regress by about 15%. And obviously there will be some memory overhead
> > as well. That's the reason we don't enable them in Android by
> > default."
>
> This should be reported and investigated. Because per-application memcg
> vs. memcg in general shouldn't make much of a difference from the
> performance side. I can see a potential performance impact for no-memcg
> vs. memcg case but even then 15% is quite a lot.
Less efficiency on memory reclaim caused by multi-LRU should be one of
the reason, which has been proved by comparing per-app memcg on/off.
Besides, theoretically workingset could also broken as LRU is too
short to compose workingset.
>
> > > My very vague understanding is that the Android system would like to
> > > freeze specific applications and for that it requires each application
> > > to live in its own cgroup. This clashes with a requirement to age and
> > > reclaim memory on a different granularity (aka no per process reclaim).
> > > So in fact something that cgroup v1 would achieve by having 2
> > > hierarchies, one for the freezer which would have a dedicated cgroup for
> > > each application and the other for the memory controller where tasks are
> > > grouped by a different criteria. This would rule out that a global (or
> > > any external memory pressure) reclaim would age LRUs that contain a mix
> > > bag of application pages rather than iterate over per-application LRUs.
> > > Is that understanding correct?
> > Correct, this is just our confusion. Besides, we believe that charge
> > the pages to implicit memory enabled parent control group doesn't make
> > sense as the memory cannot be managed at all.
>
> I do not get that part. The parent can manange and control the memory
> usage so how come it cannot be managed at all?
What I mean is the kind of parent which is enabled implicitly by
enabling on its sibling group like belowing hierarchy. Imagine that C
has no intention of memory control but has to be enabled as B would
have it. IMO, it doesn't make sense to charge C1's memory.current to C
until an explicitly echo "+memory" >  C/subtree_control.
A----B---B1
     \ C---C1
> --
> Michal Hocko
> SUSE Labs