From: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
To: Michal Hocko <mhocko@kernel.org>
Cc: "linux-mm@kvack.org" <linux-mm@kvack.org>,
l.roehrs@profihost.ag, cgroups@vger.kernel.org,
Johannes Weiner <hannes@cmpxchg.org>,
Vlastimil Babka <vbabka@suse.cz>
Subject: Re: lot of MemAvailable but falling cache and raising PSI
Date: Tue, 10 Sep 2019 11:37:19 +0200 [thread overview]
Message-ID: <743a047e-a46f-32fa-1fe4-a9bd8f09ed87@profihost.ag> (raw)
In-Reply-To: <20190910090241.GM2063@dhcp22.suse.cz>
Am 10.09.19 um 11:02 schrieb Michal Hocko:
> On Tue 10-09-19 10:38:25, Stefan Priebe - Profihost AG wrote:
>> Am 10.09.19 um 10:29 schrieb Michal Hocko:
>>> On Tue 10-09-19 07:56:36, Stefan Priebe - Profihost AG wrote:
>>>>
>>>> Am 09.09.19 um 14:56 schrieb Stefan Priebe - Profihost AG:
>>>>> Am 09.09.19 um 14:49 schrieb Michal Hocko:
>>>>>> On Mon 09-09-19 14:37:52, Stefan Priebe - Profihost AG wrote:
>>>>>>>
>>>>>>> Am 09.09.19 um 14:28 schrieb Michal Hocko:
>>>>>>>> On Mon 09-09-19 14:10:02, Stefan Priebe - Profihost AG wrote:
>>>>>>>>>
>>>>>>>>> Am 09.09.19 um 14:08 schrieb Michal Hocko:
>>>>>>>>>> On Mon 09-09-19 13:01:36, Michal Hocko wrote:
>>>>>>>>>>> and that matches moments when we reclaimed memory. There seems to be a
>>>>>>>>>>> steady THP allocations flow so maybe this is a source of the direct
>>>>>>>>>>> reclaim?
>>>>>>>>>>
>>>>>>>>>> I was thinking about this some more and THP being a source of reclaim
>>>>>>>>>> sounds quite unlikely. At least in a default configuration because we
>>>>>>>>>> shouldn't do anything expensinve in the #PF path. But there might be a
>>>>>>>>>> difference source of high order (!costly) allocations. Could you check
>>>>>>>>>> how many allocation requests like that you have on your system?
>>>>>>>>>>
>>>>>>>>>> mount -t debugfs none /debug
>>>>>>>>>> echo "order > 0" > /debug/tracing/events/kmem/mm_page_alloc/filter
>>>>>>>>>> echo 1 > /debug/tracing/events/kmem/mm_page_alloc/enable
>>>>>>>>>> cat /debug/tracing/trace_pipe > $file
>>>>>>>>
>>>>>>>> echo 1 > /debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_begin/enable
>>>>>>>> echo 1 > /debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_end/enable
>>>>>>>>
>>>>>>>> might tell us something as well but it might turn out that it just still
>>>>>>>> doesn't give us the full picture and we might need
>>>>>>>> echo stacktrace > /debug/tracing/trace_options
>>>>>>>>
>>>>>>>> It will generate much more output though.
>>>>>>>>
>>>>>>>>> Just now or when PSI raises?
>>>>>>>>
>>>>>>>> When the excessive reclaim is happening ideally.
>>>>>>>
>>>>>>> This one is from a server with 28G memfree but memory pressure is still
>>>>>>> jumping between 0 and 10%.
>>>>>>>
>>>>>>> I did:
>>>>>>> echo "order > 0" >
>>>>>>> /sys/kernel/debug/tracing/events/kmem/mm_page_alloc/filter
>>>>>>>
>>>>>>> echo 1 > /sys/kernel/debug/tracing/events/kmem/mm_page_alloc/enable
>>>>>>>
>>>>>>> echo 1 >
>>>>>>> /sys/kernel/debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_begin/enable
>>>>>>>
>>>>>>> echo 1 >
>>>>>>> /sys/kernel/debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_end/enable
>>>>>>>
>>>>>>> timeout 120 cat /sys/kernel/debug/tracing/trace_pipe > /trace
>>>>>>>
>>>>>>> File attached.
>>>>>>
>>>>>> There is no reclaim captured in this trace dump.
>>>>>> $ zcat trace1.gz | sed 's@.*\(order=[0-9]\).*\(gfp_flags=.*\)@\1 \2@' | sort | uniq -c
>>>>>> 777 order=1 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 663 order=1 gfp_flags=__GFP_IO|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 153 order=1 gfp_flags=__GFP_IO|__GFP_NOWARN|__GFP_RETRY_MAYFAIL|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 911 order=1 gfp_flags=GFP_KERNEL_ACCOUNT|__GFP_ZERO
>>>>>> 4872 order=1 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_ACCOUNT
>>>>>> 62 order=1 gfp_flags=GFP_NOWAIT|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 14 order=2 gfp_flags=GFP_ATOMIC|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP
>>>>>> 11 order=2 gfp_flags=GFP_ATOMIC|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_RECLAIMABLE
>>>>>> 1263 order=2 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 45 order=2 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC|__GFP_RECLAIMABLE
>>>>>> 1 order=2 gfp_flags=GFP_KERNEL|__GFP_COMP|__GFP_ZERO
>>>>>> 7853 order=2 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_ACCOUNT
>>>>>> 73 order=3 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 729 order=3 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC|__GFP_RECLAIMABLE
>>>>>> 528 order=3 gfp_flags=__GFP_IO|__GFP_NOWARN|__GFP_RETRY_MAYFAIL|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 1203 order=3 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_ACCOUNT
>>>>>> 5295 order=3 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP
>>>>>> 1 order=3 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 132 order=3 gfp_flags=GFP_NOWAIT|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC
>>>>>> 13 order=5 gfp_flags=GFP_KERNEL|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_ZERO
>>>>>> 1 order=6 gfp_flags=GFP_KERNEL|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_ZERO
>>>>>> 1232 order=9 gfp_flags=GFP_TRANSHUGE
>>>>>> 108 order=9 gfp_flags=GFP_TRANSHUGE|__GFP_THISNODE
>>>>>> 362 order=9 gfp_flags=GFP_TRANSHUGE_LIGHT|__GFP_THISNODE
>>>>>>
>>>>>> Nothing really stands out because except for the THP ones none of others
>>>>>> are going to even be using movable zone.
>>>>> It might be that this is not an ideal example is was just the fastest i
>>>>> could find. May be we really need one with much higher pressure.
>>>>
>>>> here another trace log where a system has 30GB free memory but is under
>>>> constant pressure and does not build up any file cache caused by memory
>>>> pressure.
>>>
>>> So the reclaim is clearly induced by THP allocations
>>> $ zgrep vmscan trace2.gz | grep gfp_flags | sed 's@.*\(gfp_flags=.*\) .*@\1@' | sort | uniq -c
>>> 1580 gfp_flags=GFP_TRANSHUGE
>>> 15 gfp_flags=GFP_TRANSHUGE|__GFP_THISNODE
>>>
>>> $ zgrep vmscan trace2.gz | grep nr_reclaimed | sed 's@nr_reclaimed=@@' | awk '{nr+=$6+0}END{print nr}'
>>> 1541726
>>>
>>> 6GB of memory reclaimed in 1776s. That is a lot! But the THP allocation
>>> rate is really high as well
>>> $ zgrep "page_alloc.*GFP_TRANSHUGE" trace2.gz | wc -l
>>> 15340
>>>
>>> this is 30GB worth of THPs (some of them might get released of course).
>>> Also only 10% of requests ends up reclaiming.
>>>
>>> One additional interesting point
>>> $ zgrep vmscan trace2.gz | grep nr_reclaimed | sed 's@.*nr_reclaimed=\([[0-9]*\)@\1@' | calc_min_max.awk
>>> min: 1.00 max: 2792.00 avg: 965.99 std: 331.12 nr: 1596
>>>
>>> Even though the std is high there are quite some outliers when a lot of
>>> memory is reclaimed.
>>>
>>> Which kernel version is this. And again, what is the THP configuration.
>>
>> This is 4.19.66 regarding THP you mean this:
>
> Do you see the same behavior with 5.3?
I rebootet with 5.3.0-rc8 - let's see what happens it might take some
hours or even days.
>> /sys/kernel/mm/transparent_hugepage/defrag:always defer [defer+madvise]
>> madvise never
>>
>> /sys/kernel/mm/transparent_hugepage/enabled:[always] madvise never
>>
>> /sys/kernel/mm/transparent_hugepage/hpage_pmd_size:2097152
>>
>> /sys/kernel/mm/transparent_hugepage/shmem_enabled:always within_size
>> advise [never] deny force
>>
>> /sys/kernel/mm/transparent_hugepage/use_zero_page:1
>>
>> /sys/kernel/mm/transparent_hugepage/enabled was madvise until yesterday
>> where i tried to switch to defer+madvise - which didn't help.
>
> Many processes hitting the reclaim are php5 others I cannot say because
> their cmd is not reflected in the trace. I suspect those are using
> madvise. I haven't really seen kcompactd interfering much. That would
> suggest using defer.
You mean i should set transparent_hugepage to defer?
Stefan
next prev parent reply other threads:[~2019-09-10 9:37 UTC|newest]
Thread overview: 61+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-09-05 11:27 lot of MemAvailable but falling cache and raising PSI Stefan Priebe - Profihost AG
2019-09-05 11:40 ` Michal Hocko
2019-09-05 11:56 ` Stefan Priebe - Profihost AG
2019-09-05 16:28 ` Yang Shi
2019-09-05 17:26 ` Stefan Priebe - Profihost AG
2019-09-05 18:46 ` Yang Shi
2019-09-05 19:31 ` Stefan Priebe - Profihost AG
2019-09-06 10:08 ` Stefan Priebe - Profihost AG
2019-09-06 10:25 ` Vlastimil Babka
2019-09-06 18:52 ` Yang Shi
2019-09-07 7:32 ` Stefan Priebe - Profihost AG
2019-09-09 8:27 ` Michal Hocko
2019-09-09 8:54 ` Stefan Priebe - Profihost AG
2019-09-09 11:01 ` Michal Hocko
2019-09-09 12:08 ` Michal Hocko
2019-09-09 12:10 ` Stefan Priebe - Profihost AG
2019-09-09 12:28 ` Michal Hocko
2019-09-09 12:37 ` Stefan Priebe - Profihost AG
2019-09-09 12:49 ` Michal Hocko
2019-09-09 12:56 ` Stefan Priebe - Profihost AG
[not found] ` <52235eda-ffe2-721c-7ad7-575048e2d29d@profihost.ag>
2019-09-10 5:58 ` Stefan Priebe - Profihost AG
2019-09-10 8:29 ` Michal Hocko
2019-09-10 8:38 ` Stefan Priebe - Profihost AG
2019-09-10 9:02 ` Michal Hocko
2019-09-10 9:37 ` Stefan Priebe - Profihost AG [this message]
2019-09-10 11:07 ` Michal Hocko
2019-09-10 12:45 ` Stefan Priebe - Profihost AG
2019-09-10 12:57 ` Michal Hocko
2019-09-10 13:05 ` Stefan Priebe - Profihost AG
2019-09-10 13:14 ` Stefan Priebe - Profihost AG
2019-09-10 13:24 ` Michal Hocko
2019-09-11 6:12 ` Stefan Priebe - Profihost AG
2019-09-11 6:24 ` Stefan Priebe - Profihost AG
2019-09-11 13:59 ` Stefan Priebe - Profihost AG
2019-09-12 10:53 ` Stefan Priebe - Profihost AG
2019-09-12 11:06 ` Stefan Priebe - Profihost AG
2019-09-11 7:09 ` 5.3-rc-8 hung task in IO (was: Re: lot of MemAvailable but falling cache and raising PSI) Michal Hocko
2019-09-11 14:09 ` Stefan Priebe - Profihost AG
2019-09-11 14:56 ` Filipe Manana
2019-09-11 15:39 ` Stefan Priebe - Profihost AG
2019-09-11 15:56 ` Filipe Manana
2019-09-11 16:15 ` Stefan Priebe - Profihost AG
2019-09-11 16:19 ` Filipe Manana
2019-09-19 10:21 ` lot of MemAvailable but falling cache and raising PSI Stefan Priebe - Profihost AG
2019-09-23 12:08 ` Michal Hocko
2019-09-27 12:45 ` Vlastimil Babka
2019-09-30 6:56 ` Stefan Priebe - Profihost AG
2019-09-30 7:21 ` Vlastimil Babka
2019-10-22 7:41 ` Stefan Priebe - Profihost AG
2019-10-22 7:48 ` Vlastimil Babka
2019-10-22 10:02 ` Stefan Priebe - Profihost AG
2019-10-22 10:20 ` Oscar Salvador
2019-10-22 10:21 ` Vlastimil Babka
2019-10-22 11:08 ` Stefan Priebe - Profihost AG
2019-09-10 5:41 ` Stefan Priebe - Profihost AG
2019-09-09 11:49 ` Vlastimil Babka
2019-09-09 12:09 ` Stefan Priebe - Profihost AG
2019-09-09 12:21 ` Vlastimil Babka
2019-09-09 12:31 ` Stefan Priebe - Profihost AG
2019-09-05 12:15 ` Vlastimil Babka
2019-09-05 12:27 ` Stefan Priebe - Profihost AG
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=743a047e-a46f-32fa-1fe4-a9bd8f09ed87@profihost.ag \
--to=s.priebe@profihost.ag \
--cc=cgroups@vger.kernel.org \
--cc=hannes@cmpxchg.org \
--cc=l.roehrs@profihost.ag \
--cc=linux-mm@kvack.org \
--cc=mhocko@kernel.org \
--cc=vbabka@suse.cz \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).