Re: lot of MemAvailable but falling cache and raising PSI - Stefan Priebe

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: Wed, 11 Sep 2019 08:24:25 +0200	[thread overview]
Message-ID: <5f960e74-1f44-9a0a-58a6-dcb64aa71612@profihost.ag> (raw)
In-Reply-To: <d07620d9-4967-40fe-fa0f-be51f2459dc5@profihost.ag>

Hi Michal,

Am 11.09.19 um 08:12 schrieb Stefan Priebe - Profihost AG:
> Hi Michal,
> Am 10.09.19 um 15:24 schrieb Michal Hocko:
>> On Tue 10-09-19 15:14:45, Stefan Priebe - Profihost AG wrote:
>>> Am 10.09.19 um 15:05 schrieb Stefan Priebe - Profihost AG:
>>>>
>>>> Am 10.09.19 um 14:57 schrieb Michal Hocko:
>>>>> On Tue 10-09-19 14:45:37, Stefan Priebe - Profihost AG wrote:
>>>>>> Hello Michal,
>>>>>>
>>>>>> ok this might take a long time. Attached you'll find a graph from a
>>>>>> fresh boot what happens over time (here 17 August to 30 August). Memory
>>>>>> Usage decreases as well as cache but slowly and only over time and days.
>>>>>>
>>>>>> So it might take 2-3 weeks running Kernel 5.3 to see what happens.
>>>>>
>>>>> No problem. Just make sure to collect the requested data from the time
>>>>> you see the actual problem. Btw. you try my very dumb scriplets to get
>>>>> an idea of how much memory gets reclaimed due to THP.
>>>>
>>>> You mean your sed and sort on top of the trace file? No i did not with
>>>> the current 5.3 kernel do you think it will show anything interesting?
>>>> Which line shows me how much memory gets reclaimed due to THP?
>>
>> Please re-read http://lkml.kernel.org/r/20190910082919.GL2063@dhcp22.suse.cz
>> Each command has a commented output. If you see nunmber of reclaimed
>> pages to be large for GFP_TRANSHUGE then you are seeing a similar
>> problem.
>>
>>> Is something like a kernel memory leak possible? Or wouldn't this end up
>>> in having a lot of free memory which doesn't seem usable.
>>
>> I would be really surprised if this was the case.
>>
>>> I also wonder why a reclaim takes place when there is enough memory.
>>
>> This is not clear yet and it might be a bug that has been fixed since
>> 4.18. That's why we need to see whether the same is pattern is happening
>> with 5.3 as well.

but except from the btrfs problem the memory consumption looks far
better than before.

Running 4.19.X:
after about 12h cache starts to drop from 30G to 24G

Running 5.3-rc8:
after about 24h cache is still constant at nearly 30G

Greets,
Stefan