Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <1660908562-17409-1-git-send-email-zhaoyang.huang@unisoc.com>
 <Yv+6YjaGAv52yvq9@slm.duckdns.org> <CALvZod7QdLSMdBoD2WztL72qS8kJe7F79JuCH6t19rRcw6Pn1w@mail.gmail.com>
 <Yv/EArPDTcCrGqJh@slm.duckdns.org> <YwNpI1ydy0yDnBH0@dhcp22.suse.cz>
 <CAGWkznEB+R0YBaBFBL7dPqs8R=qKC6+ixTWEGCYy2PaczXkaPA@mail.gmail.com>
 <YwRjyx6wFLk8WTDe@dhcp22.suse.cz> <CAGWkznGaYTv4u4kOo-rupfyWzDNJXNKTchwP6dbUK-=UXWm47w@mail.gmail.com>
 <YwSQ4APOu/H7lYGL@dhcp22.suse.cz> <CAGWkznGd6mgareABseMKY5p0f1=5dkfVkj=NS7_B6OkXBYSwyw@mail.gmail.com>
 <YwS/S9Sd1OWnT81Q@dhcp22.suse.cz>
In-Reply-To: <YwS/S9Sd1OWnT81Q@dhcp22.suse.cz>
From:   Suren Baghdasaryan <surenb@google.com>
Date:   Tue, 23 Aug 2022 09:21:16 -0700
Message-ID: <CAJuCfpH+T9+eOVYDfOv9yNSfAvq=pJtMp4ZaoYaM7iMY9XkaUw@mail.gmail.com>
Subject: Re: [RFC PATCH] memcg: use root_mem_cgroup when css is inherited
To:     Michal Hocko <mhocko@suse.com>
Cc:     Zhaoyang Huang <huangzhaoyang@gmail.com>,
        Tejun Heo <tj@kernel.org>, Shakeel Butt <shakeelb@google.com>,
        "zhaoyang.huang" <zhaoyang.huang@unisoc.com>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Linux MM <linux-mm@kvack.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Cgroups <cgroups@vger.kernel.org>, Ke Wang <ke.wang@unisoc.com>,
        Zefan Li <lizefan.x@bytedance.com>,
        Roman Gushchin <roman.gushchin@linux.dev>,
        Muchun Song <songmuchun@bytedance.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Aug 23, 2022 at 4:51 AM Michal Hocko <mhocko@suse.com> wrote:
>
> On Tue 23-08-22 17:20:59, Zhaoyang Huang wrote:
> > On Tue, Aug 23, 2022 at 4:33 PM Michal Hocko <mhocko@suse.com> wrote:
> > >
> > > On Tue 23-08-22 14:03:04, Zhaoyang Huang wrote:
> > > > On Tue, Aug 23, 2022 at 1:21 PM Michal Hocko <mhocko@suse.com> wrote:
> > > > >
> > > > > On Tue 23-08-22 10:31:57, Zhaoyang Huang wrote:
> > > [...]
> > > > > > 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.
> > >
> > > Do you have any data to back these claims? Is this something that could
> > > be handled on the configuration level? E.g. by applying low limit
> > > protection to keep the workingset in the memory?
> > I don't think so. IMO, workingset works when there are pages evicted
> > from LRU and then refault which provide refault distance for pages.
> > Applying memcg's protection will have all LRU out of evicted which
> > make the mechanism fail.
>
> It is really hard to help you out without any actual data. The idea was
> though to use the low limit protection to adaptively configure
> respective memcgs to reduce refaults. You already have data about
> refaults ready so increasing the limit for often refaulting memcgs would
> reduce the trashing.

Sorry for joining late.
A couple years ago I tested root-memcg vs per-app memcg configurations
on an Android phone. Here is a snapshot from my findings:

Problem
=======
We see tangible increase in major faults and workingset refaults when
transitioning from root-only memory cgroup to per-application cgroups
on Android.

Test results
============
Results while running memory-demanding workload:
root memcg     per-app memcg     delta
workingset_refault 1771228 3874281 +118.73%
workingset_nodereclaim 4543 13928 +206.58%
pgpgin 13319208 20618944 +54.81%
pgpgout 1739552 3080664 +77.1%
pgpgoutclean 2616571 4805755 +83.67%
pswpin 359211 3918716 +990.92%
pswpout 1082238 5697463 +426.45%
pgfree 28978393 32531010 +12.26%
pgactivate 2586562 8731113 +237.56%
pgdeactivate 3811074 11670051 +206.21%
pgfault 38692510 46096963 +19.14%
pgmajfault 441288 4100020 +829.1%
pgrefill 4590451 12768165 +178.15%

Results while running application cycle test (20 apps, 20 cycles):
root memcg     per-app memcg     delta
workingset_refault 10634691 11429223 +7.47%
workingset_nodereclaim 37477 59033 +57.52%
pgpgin 70662840 69569516 -1.55%
pgpgout 2605968 2695596 +3.44%
pgpgoutclean 13514955 14980610 +10.84%
pswpin 1489851 3780868 +153.77%
pswpout 4125547 8050819 +95.15%
pgfree 99823083 105104637 +5.29%
pgactivate 7685275 11647913 +51.56%
pgdeactivate 14193660 21459784 +51.19%
pgfault 89173166 100598528 +12.81%
pgmajfault 1856172 4227190 +127.74%
pgrefill 16643554 23203927 +39.42%

Tests were conducted on an Android phone with 4GB RAM.
Similar regression was reported a couple years ago here:
https://www.spinics.net/lists/linux-mm/msg121665.html

I plan on checking the difference again on newer kernels (likely 5.15)
after LPC this September.

>
> [...]
> > > A.cgroup.controllers = memory
> > > A.cgroup.subtree_control = memory
> > >
> > > A/B.cgroup.controllers = memory
> > > A/B.cgroup.subtree_control = memory
> > > A/B/B1.cgroup.controllers = memory
> > >
> > > A/C.cgroup.controllers = memory
> > > A/C.cgroup.subtree_control = ""
> > > A/C/C1.cgroup.controllers = ""
> > Yes for above hierarchy and configuration.
> > >
> > > Is your concern that C1 is charged to A/C or that you cannot actually make
> > > A/C.cgroup.controllers = "" because you want to maintain memory in A?
> > > Because that would be breaking the internal node constrain rule AFAICS.
> > No. I just want to keep memory on B.
>
> That would require A to be without controllers which is not possible due
> to hierarchical constrain.
>
> > > Or maybe you just really want a different hierarchy where
> > > A == root_cgroup and want the memory acocunted in B
> > > (root/B.cgroup.controllers = memory) but not in C (root/C.cgroup.controllers = "")?
> > Yes.
> > >
> > > That would mean that C memory would be maintained on the global (root
> > > memcg) LRUs which is the only internal node which is allowed to have
> > > resources because it is special.
> > Exactly. I would like to have all groups like C which have no parent's
> > subtree_control = memory charge memory to root. Under this
> > implementation, memory under enabled group will be protected by
> > min/low while other groups' memory share the same LRU to have
> > workingset things take effect.
>
> One way to achieve that would be shaping the hierarchy the following way
>             root
>         /         \
> no_memcg[1]      memcg[2]
> ||||||||         |||||
> app_cgroups     app_cgroups
>
> with
> no_memcg.subtree_control = ""
> memcg.subtree_control = memory
>
> no?
>
> You haven't really described why you need per application freezer cgroup
> but I suspect you want to selectively freeze applications. Is there
> any obstacle to have a dedicated frozen cgroup and migrate tasks to be
> frozen there?

We intend for Android to gradually migrate to v2 cgroups for all
controllers and given that it has to use a unified hierarchy,
per-application hierarchy provides highest flexibility. That way we
can control every aspect of every app without affecting others. Of
course that comes with its overhead.
Thanks,
Suren.

> --
> Michal Hocko
> SUSE Labs