Re: PROBLEM: cgroup cost too much memory when transfer small files to tmpfs

[Shakeel Butt <shakeelb@google.com>]

On Tue, Jul 21, 2020 at 11:51 AM Roman Gushchin <guro@fb.com> wrote:
>
> On Tue, Jul 21, 2020 at 01:41:26PM -0400, Johannes Weiner wrote:
> > On Tue, Jul 21, 2020 at 11:19:52AM +0000, jingrui wrote:
> > > Cc: Johannes Weiner <hannes@cmpxchg.org> ; Michal Hocko <mhocko@kernel.org>; Vladimir Davydov <vdavydov.dev@gmail.com>
> > >
> > > Thanks.
> > >
> > > ---
> > > PROBLEM: cgroup cost too much memory when transfer small files to tmpfs.
> > >
> > > keywords: cgroup PERCPU/memory cost too much.
> > >
> > > description:
> > >
> > > We send small files from node-A to node-B tmpfs /tmp directory using sftp. On
> > > node-B the systemd configured with pam on like below.
> > >
> > > cat /etc/pam.d/password-auth | grep systemd
> > > -session     optional      pam_systemd.so
> > >
> > > So when transfer a file, a systemd session is created, that means a cgroup is
> > > created, then file saved at /tmp will associated with a cgroup object. After
> > > file transferred, session and cgroup-dir will be removed, but the file in /tmp
> > > still associated with the cgroup object. The PERCPU memory in cgroup/css object
> > > cost a lot(about 0.5MB/per-cgroup-object) on 200/cpus machine.
> >
> > CC Roman who had a patch series to free all this extended (percpu)
> > memory upon cgroup deletion:
> >
> > https://lore.kernel.org/patchwork/cover/1050508/
> >
> > It looks like it never got merged for some reason.
>
> The mentioned patchset can make the problem less noticeable, but can't solve it completely.
> It has never been merged, because the dying cgroup problem was mostly solved by other methods:
> slab memory reparenting and various reclaim fixes. So there was no more reason to complicate
> the code to release the memcg memory early.
>
> The overhead of creating and destroying a new memory cgroup for a transfer of a small
> file will be noticeable anyway. So IMO the solution is to use a single cgroup for all
> transfers. I don't know if systemd supports such mode out of the box, but it shouldn't
> be hard to add it.
>
> But also I wonder if we need a special tmpfs mount option, something like "noaccount".
> Not only for this specific case, but also for the case when tmpfs is extensively
> shared between multiple cgroups or if it's used to pass some data from one cgroup
> to another, or if we care about the performance more than about the accounting;
> in other words for cases where the accounting makes more harm than good.
>

Internally we actually have an tmpfs mount option "memcg=" which
charges all the memory of the tmpfs files on that mount to the given
memcg and the motivation is the shared tmpfs files between multiple
cgroups. One concrete use-case is the shared memory used for
communication between the application and the user space network
driver [1]. The "memcg=root" can be used as a "noaccount" option.

[1] https://sosp19.rcs.uwaterloo.ca/slides/marty.pdf