Re: [PATCH 00/14] per memcg lru_lock

From: Konstantin Khlebnikov
Date: Mon Aug 26 2019 - 04:39:56 EST


On 22/08/2019 18.20, Daniel Jordan wrote:
On 8/22/19 7:56 AM, Alex Shi wrote:
å 2019/8/22 äå2:00, Daniel Jordan åé:
ÂÂ https://git.kernel.org/pub/scm/linux/kernel/git/wfg/vm-scalability.git/tree/case-lru-file-readtwice>
It's also synthetic but it stresses lru_lock more than just anon alloc/free. It hits the page activate path, which is where we see this lock in our database, and if enough memory is configured lru_lock also gets stressed during reclaim, similar to [1].

Thanks for the sharing, this patchset can not help the [1] case, since it's just relief the per container lock contention now.

I should've been clearer. [1] is meant as an example of someone suffering from lru_lock during reclaim. Wouldn't your series help per-memcg reclaim?

Yes, readtwice case could be more sensitive for this lru_lock changes in containers. I may try to use it in container with some tuning. But anyway, aim9 is also pretty good to show the problem and solutions. :)

It'd be better though, as Michal suggests, to use the real workload that's causing problems. Where are you seeing contention?

We repeatly create or delete a lot of different containers according to servers load/usage, so normal workload could cause lots of pages alloc/remove.

I think numbers from that scenario would help your case.

aim9 could reflect part of scenarios. I don't know the DB scenario yet.

We see it during DB shutdown when each DB process frees its memory (zap_pte_range -> mark_page_accessed). But that's a different thing, clearly Not This Series.

With this patch series, lruvec->lru_lock show no contentions
ÂÂÂÂÂÂÂÂÂ &(&lruvec->lru_l...ÂÂÂÂÂÂÂÂÂ 8ÂÂÂÂÂÂÂÂÂ 0ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0ÂÂÂÂÂÂ 0ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0

and aim9 page_test/brk_test performance increased 5%~50%.

Where does the 50% number come in? The numbers below seem to only show ~4% boost.
After splitting lru-locks present per-cpu page-vectors works no so well
because they mixes pages from different cgroups.

pagevec_lru_move_fn and friends need better implementation:
either sorting pages or splitting vectores in per-lruvec basis.
the Setddev/CoeffVar case has about 50% performance increase. one of container's mmtests result as following:

StddevÂÂÂ page_testÂÂÂÂÂ 245.15 (ÂÂ 0.00%)ÂÂÂÂÂ 189.29 (Â 22.79%)
StddevÂÂÂ brk_testÂÂÂÂÂ 1258.60 (ÂÂ 0.00%)ÂÂÂÂÂ 629.16 (Â 50.01%)
CoeffVar page_test 0.71 ( 0.00%) 0.53 ( 26.05%)
CoeffVar brk_test 1.32 ( 0.00%) 0.64 ( 51.14%)

Aha. 50% decrease in stdev.


After splitting lru-locks present per-cpu page-vectors works
no so well because they mix pages from different cgroups.

pagevec_lru_move_fn and friends need better implementation:
either sorting pages or splitting vectores in per-lruvec basis.