Re: [PATCH RFC 0/3] mm: Reduce IO by improving algorithm of memcg pagecache pages eviction
From: Kirill Tkhai
Date: Wed Jan 09 2019 - 10:43:16 EST
Hi, Michal,
On 09.01.2019 17:11, Michal Hocko wrote:
> On Wed 09-01-19 15:20:18, Kirill Tkhai wrote:
>> On nodes without memory overcommit, it's common a situation,
>> when memcg exceeds its limit and pages from pagecache are
>> shrinked on reclaim, while node has a lot of free memory.
>
> Yes, that is the semantic of the hard limit. If the system is not
> overcommitted then the hard limit can be used to prevent unexpected
> direct reclaim from unrelated activity.
According to Documentation/admin-guide/cgroup-v2.rst:
memory.max
Memory usage hard limit. This is the final protection
mechanism. If a cgroup's memory usage reaches this limit and
can't be reduced, the OOM killer is invoked in the cgroup.
Under certain circumstances, the usage may go over the limit
temporarily.
There is nothing about direct reclaim in another memcg. I don't think
we break something here.
File pages are accounted to memcg, and this guarantees, that single
memcg won't occupy all system memory by its unevictible page cache.
But the suggested patchset follows the same way. Pages, which remain
in pagecache, are easy-to-be-evicted, since they are not dirty and
not under writeback. System can drop them fast and in foreseeable time.
This is cardinal thing about the patchset: remained pages do not
introduce principal burden on system memory or reclaim time.
>> Further access to the pages requires real device IO, while
>> IO causes time delays, worse powerusage, worse throughput
>> for other users of the device, etc.
>
> It is to be expected that a memory throttled usage will have this side
> effect IMO.
>
>> Cleancache is not a good solution for this problem, since
>> it implies copying of page on every cleancache_put_page()
>> and cleancache_get_page(). Also, it requires introduction
>> of internal per-cleancache_ops data structures to manage
>> cached pages and their inodes relationships, which again
>> introduces overhead.
>>
>> This patchset introduces another solution. It introduces
>> a new scheme for evicting memcg pages:
>>
>> 1)__remove_mapping() uncharges unmapped page memcg
>> and leaves page in pagecache on memcg reclaim;
>>
>> 2)putback_lru_page() places page into root_mem_cgroup
>> list, since its memcg is NULL. Page may be evicted
>> on global reclaim (and this will be easily, as
>> page is not mapped, so shrinker will shrink it
>> with 100% probability of success);
>>
>> 3)pagecache_get_page() charges page into memcg of
>> a task, which takes it first.
>
> But this also means that any hard limited memcg can fill up all the
> memory and break the above assumption about the isolation from direct
> reclaim. Not to mention the OOM or is there anything you do anything
> about preventing that?
This is discussed thing. We may add such the pages into tail of LRU list
instead of head. We may introduce one more separate list to link such
the pages only, and fastly evict them in case of global reclaim. I don't
think there is a problem.
> That beig said, I do not think we want to or even can change the
> semantic of the hard limit and break existing setups.
Using the original description and the comments I gave in this message,
could you please to clarify the way we break existing setups?
> I am still
> interested to hear more about more detailed/specific usecases that might
> benefit from this behavior. Why do those users even use hard limit at
> all? To protect from anon memory leaks?
In multi-user machine people want to have size of available to container
memory equal to the size, which they pay. So, hard limit is needed to prevent
one container to occupy all system memory via slowly-evictible writeback
pages, unevictible anon pages, etc. You can't fastly allocate a page,
in case of many pages are under writeback, this operation is very slow.
(But unmapped pagecache pages introduced by patchset is another thing:
you just need to take not sleeping spinlock to call __delete_from_page_cache()
only. This is fast)
Multi-user machine may have more memory, than sum of all containers hard
limit. This may be used as an optimization just to reduce disk IO. There
is no contradiction to sane sense here. And it's not a rare situation.
In our kernel we have cleancache driver for handling this situation, but
cleancache is not the best solution like I wrote.
Not overcommited system is likely case for the patchset, while the below
is a little less likely:
> Do different memcgs share the page cache heavily?
People may use NFS server from different containers. They may want
to have low priority userspace workers and high priority main task,
and they may want to reduce the workers memory consumption. These
are the cases.
Kirill