On 4/21/22 12:33, Roman Gushchin wrote:
On Thu, Apr 21, 2022 at 10:58:45AM -0400, Waiman Long wrote:
For systems with large number of CPUs, the majority of the memory
consumed by the mem_cgroup structure is actually the percpu stats
memory. When a large number of memory cgroups are continuously created
and destroyed (like in a container host), it is possible that more
and more mem_cgroup structures remained in the dying state holding up
increasing amount of percpu memory.
We can't free up the memory of the dying mem_cgroup structure due to
active references in some other places. However, the percpu stats memory
allocated to that mem_cgroup is a different story.
This patch adds a new percpu_stats_disabled variable to keep track of
the state of the percpu stats memory. If the variable is set, percpu
stats update will be disabled for that particular memcg. All the stats
update will be forward to its parent instead. Reading of the its percpu
stats will return 0.
The flushing and freeing of the percpu stats memory is a multi-step
process. The percpu_stats_disabled variable is set when the memcg is
being set to offline state. After a grace period with the help of RCU,
the percpu stats data are flushed and then freed.
This will greatly reduce the amount of memory held up by dying memory
cgroups.
By running a simple management tool for container 2000 times per test
run, below are the results of increases of percpu memory (as reported
in /proc/meminfo) and nr_dying_descendants in root's cgroup.stat.
Hi Waiman!
I've been proposing the same idea some time ago:
https://lore.kernel.org/all/20190312223404.28665-7-guro@xxxxxx/T/ .
However I dropped it with the thinking that with many other fixes
preventing the accumulation of the dying cgroups it's not worth the added
complexity and a potential cpu overhead.
I think it ultimately comes to the number of dying cgroups. If it's low,
memory savings are not worth the cpu overhead. If it's high, they are.
I hope long-term to drive it down significantly (with lru-pages reparenting
being the first major milestone), but it might take a while.
I don't have a strong opinion either way, just want to dump my thoughts
on this.
I have quite a number of customer cases complaining about increasing
percpu memory usages. The number of dying memcg's can go to tens of
thousands. From my own investigation, I believe that those dying memcg's
are not freed because they are pinned down by references in the page
structure. I am aware that we support the use of objcg in the page
structure which will allow easy reparenting, but most pages don't do
that and it is not easy to do this conversion and it may take quite a
while to do that.
In term of overhead, it is mostly one more memory read from the
mem_cgroup structure in the update path. I don't expect there will be
that many updates when the memcg is in an offline state as updates will
be slower in this case. Freeing the dying memcg will take a bit longer
though, but its impact on the overall system performance should still be
negligible.
I am also thinking about using a static_key for turning it on only for
systems with more than, say, 20 cpus as the percpu memory overhead
increases linearly with the number of possible cpus.
Any other suggestions and improvements are welcome.
Cheers,
Longman
