Hi Michal,
Let me try to pitch the motivation first:
In the case of NUMA-aware shrinker, memory pressure may lead to splitting and freeing subpages within a THP, irrespective of whether the page belongs to the memcg that is under memory pressure. THP sharing between memcgs is not a pre-condition for above to happen.Let's consider two memcgs: memcg-A and memcg-B. Say memcg-A is under memory pressure that is hitting its limit. If this memory pressure invokes the shrinker (non-memcg-aware) and splits pages from memcg-B queued for deferred splits, then that won't reduce memcg-A's usage. It will reduce memcg-B's usage. Also, why should memcg-A's memory pressure reduce memcg-B's usage.
By making this shrinker memcg-aware, we can invoke respective memcg shrinkers to handle the memory pressure. Furthermore, with this approach we can isolate the THPs of other memcg(s) (not under memory pressure) from premature splits. Isolation aids in reducing performance impact when we have several memcgs on the same machine.
By making this shrinker memcg-aware, we can invoke respective memcg shrinkers to handle the memory pressure. Furthermore, with this approach we can isolate the THPs of other memcg(s) (not under memory pressure) from premature splits. Isolation aids in reducing performance impact when we have several memcgs on the same machine.
Regarding ifdef ugliness: I get your point and agree with you on that. I think I can do a better job at restricting the ugliness, will post another version.
Thanks,
-Neha Agarwal
On Fri, Oct 20, 2017 at 12:12 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
On Thu 19-10-17 13:03:23, Neha Agarwal wrote:
> deferred_split_shrinker is NUMA aware. Making it memcg-aware if
> CONFIG_MEMCG is enabled to prevent shrinking memory of memcg(s) that are
> not under memory pressure. This change isolates memory pressure across
> memcgs from deferred_split_shrinker perspective, by not prematurely
> splitting huge pages for the memcg that is not under memory pressure.
Why do we need this? THP pages are usually not shared between memcgs. Or
do you have a real world example where this is not the case? Your patch
is adding quite a lot of (and to be really honest very ugly) code so
there better should be a _very_ good reason to justify it. I haven't
looked very closely to the code, at least all those ifdefs in the code
are too ugly to live.
--
Michal Hocko
SUSE Labs
