On Tue, Mar 30, 2021 at 11:34:11AM -0700, Shakeel Butt wrote:
> On Tue, Mar 30, 2021 at 3:20 AM Muchun Song <songmuc...@bytedance.com> wrote:
> >
> > Since the following patchsets applied. All the kernel memory are charged
> > with the new APIs of obj_cgroup.
> >
> > [v17,00/19] The new cgroup slab memory controller
> > [v5,0/7] Use obj_cgroup APIs to charge kmem pages
> >
> > But user memory allocations (LRU pages) pinning memcgs for a long time -
> > it exists at a larger scale and is causing recurring problems in the real
> > world: page cache doesn't get reclaimed for a long time, or is used by the
> > second, third, fourth, ... instance of the same job that was restarted into
> > a new cgroup every time. Unreclaimable dying cgroups pile up, waste memory,
> > and make page reclaim very inefficient.
> >
> > We can convert LRU pages and most other raw memcg pins to the objcg
> > direction
> > to fix this problem, and then the LRU pages will not pin the memcgs.
> >
> > This patchset aims to make the LRU pages to drop the reference to memory
> > cgroup by using the APIs of obj_cgroup. Finally, we can see that the number
> > of the dying cgroups will not increase if we run the following test script.
> >
> > ```bash
> > #!/bin/bash
> >
> > cat /proc/cgroups | grep memory
> >
> > cd /sys/fs/cgroup/memory
> >
> > for i in range{1..500}
> > do
> > mkdir test
> > echo $$ > test/cgroup.procs
> > sleep 60 &
> > echo $$ > cgroup.procs
> > echo `cat test/cgroup.procs` > cgroup.procs
> > rmdir test
> > done
> >
> > cat /proc/cgroups | grep memory
> > ```
> >
> > Patch 1 aims to fix page charging in page replacement.
> > Patch 2-5 are code cleanup and simplification.
> > Patch 6-15 convert LRU pages pin to the objcg direction.
>
> The main concern I have with *just* reparenting LRU pages is that for
> the long running systems, the root memcg will become a dumping ground.
> In addition a job running multiple times on a machine will see
> inconsistent memory usage if it re-accesses the file pages which were
> reparented to the root memcg.
I don't understand how Muchun's patches are supposed to *change* the
behavior the way you are describing it. This IS today's behavior.
We have hierarchical accounting, and a page that belongs to a leaf
cgroup will automatically belong to all its parents.
Further, if you delete a cgroup today, the abandoned cache will stay
physically linked to that cgroup, but that zombie group no longer acts
as a control structure: it imposes no limit and no protection; the
pages will be reclaimed as if it WERE linked to the parent.
For all intents and purposes, when you delete a cgroup today, its
remaining pages ARE dumped onto the parent.
The only difference is that today they pointlessly pin the leaf cgroup
as a holding vessel - which is then round-robin'd from the parent
during reclaim in order to pretend that all these child pages actually
ARE linked to the parent's LRU list.
Remember how we used to have every page physically linked to multiple
lrus? The leaf cgroup and the root?
All pages always belong to the (virtual) LRU list of all ancestor
cgroups. The only thing Muchun changes is that they no longer pin a
cgroup that has no semantical meaning anymore (because it's neither
visible to the user nor exerts any contol over the pages anymore).
Maybe I'm missing something that either you or Roman can explain to
me. But this series looks like a (rare) pure win.
Whether you like the current semantics is a separate discussion IMO.
> Please note that I do agree with the mentioned problem and we do see
> this issue in our fleet. Internally we have a "memcg mount option"
> feature which couples a file system with a memcg and all file pages
> allocated on that file system will be charged to that memcg. Multiple
> instances (concurrent or subsequent) of the job will use that file
> system (with a dedicated memcg) without leaving the zombies behind. I
> am not pushing for this solution as it comes with its own intricacies
> (e.g. if memcg coupled with a file system has a limit, the oom
> behavior would be awkward and therefore internally we don't put a
> limit on such memcgs). Though I want this to be part of discussion.
Right, you disconnect memory from the tasks that are allocating it,
and so you can't assign culpability when you need to.
OOM is one thing, but there are also CPU cycles and IO bandwidth
consumed during reclaim.
> I think the underlying reasons behind this issue are:
>
> 1) Filesystem shared by disjoint jobs.
> 2) For job dedicated filesystems, the lifetime of the filesystem is
> different from the lifetime of the job.
There is also the case of deleting a cgroup just to recreate it right
after for the same job. Many job managers do this on restart right now
- like systemd, and what we're using in our fleet. This seems
avoidable by recycling a group for another instance of the same job.
Sharing is a more difficult discussion. If you access a page that you
share with another cgroup, it may or may not be subject to your own or
your buddy's memory limits. The only limit it is guaranteed to be
subjected to is that of your parent. So One thing I could imagine is,
instead of having a separate cgroup outside the hierarchy, we would
reparent live pages the second they are accessed from a foreign
cgroup. And reparent them until you reach the first common ancestor.
This way, when you mount a filesystem shared by two jobs, you can put
them into a joint subtree, and the root level of this subtree captures
all the memory (as well as the reclaim CPU and IO) used by the two
jobs - the private portions and the shared portions - and doesn't make
them the liability of jobs in the system that DON'T share the same fs.
But again, this is a useful discussion to have, but I don't quite see
why it's relevant to Muchun's patches. They're purely an optimization.
So I'd like to clear that up first before going further.
