Dirty/Writeback fields in /proc/meminfo affected by 20d74bf29c

Dirty/Writeback fields in /proc/meminfo affected by 20d74bf29c

[Tomas Vondra]

Hi,

While investigating a strange OOM issue on the 3.18.x branch (which 
turned out to be already fixed by 52c84a95), I've noticed a strange 
difference in Dirty/Writeback fields in /proc/meminfo depending on 
kernel version. I'm wondering whether this is expected ...

I've bisected the change to 20d74bf29c, added in 3.18.22 (upstream 
commit 4f258a46):

     sd: Fix maximum I/O size for BLOCK_PC requests

With /etc/sysctl.conf containing

     vm.dirty_background_bytes = 67108864
     vm.dirty_bytes = 1073741824

a simple "dd" example writing 10GB file

     dd if=/dev/zero of=ssd.test.file bs=1M count=10240

results in about this on 3.18.21:

     Dirty:            740856 kB
     Writeback:         12400 kB

but on 3.18.22:

     Dirty:             49244 kB
     Writeback:        656396 kB

I.e. it seems to revert the relationship. I haven't identified any 
performance impact, and apparently for random writes the behavior did 
not change at all (or at least I haven't managed to reproduce it).

But it's unclear to me why setting a maximum I/O size should affect 
this, and perhaps it has impact that I don't see.

regards
Tomas

Re: Dirty/Writeback fields in /proc/meminfo affected by 20d74bf29c

[Andrew Morton]

On Mon, 1 Aug 2016 04:36:28 +0200 Tomas Vondra <tomas@pgaddict.com> wrote:

> Hi,
> 
> While investigating a strange OOM issue on the 3.18.x branch (which 
> turned out to be already fixed by 52c84a95), I've noticed a strange 
> difference in Dirty/Writeback fields in /proc/meminfo depending on 
> kernel version. I'm wondering whether this is expected ...
> 
> I've bisected the change to 20d74bf29c, added in 3.18.22 (upstream 
> commit 4f258a46):
> 
>      sd: Fix maximum I/O size for BLOCK_PC requests
> 
> With /etc/sysctl.conf containing
> 
>      vm.dirty_background_bytes = 67108864
>      vm.dirty_bytes = 1073741824
> 
> a simple "dd" example writing 10GB file
> 
>      dd if=/dev/zero of=ssd.test.file bs=1M count=10240
> 
> results in about this on 3.18.21:
> 
>      Dirty:            740856 kB
>      Writeback:         12400 kB
> 
> but on 3.18.22:
> 
>      Dirty:             49244 kB
>      Writeback:        656396 kB
> 
> I.e. it seems to revert the relationship. I haven't identified any 
> performance impact, and apparently for random writes the behavior did 
> not change at all (or at least I haven't managed to reproduce it).
> 
> But it's unclear to me why setting a maximum I/O size should affect 
> this, and perhaps it has impact that I don't see.

So what appears to be happening here is that background writeback is
cutting in earlier - the amount of pending writeback ("Dirty") is
reduced while the amount of active writeback ("Writeback") is
correspondingly increased.

4f258a46 had the effect of permitting larger requests into the request
queue.  It's unclear to me why larger requests would cause background
writeback to cut in earlier - the writeback code doesn't even care
about individual request sizes, it only cares about aggregate pagecache
state.

Less Dirty and more Writeback isn't necessarily a bad thing at all, but
I don't like mysteries.  cc linux-mm to see if anyone else can
spot-the-difference.

Re: Dirty/Writeback fields in /proc/meminfo affected by 20d74bf29c

[Tomas Vondra]

On 08/04/2016 10:55 PM, Andrew Morton wrote:
> On Mon, 1 Aug 2016 04:36:28 +0200 Tomas Vondra <tomas@pgaddict.com> wrote:
>
>> Hi,
>>
>> While investigating a strange OOM issue on the 3.18.x branch (which
>> turned out to be already fixed by 52c84a95), I've noticed a strange
>> difference in Dirty/Writeback fields in /proc/meminfo depending on
>> kernel version. I'm wondering whether this is expected ...
>>
>> I've bisected the change to 20d74bf29c, added in 3.18.22 (upstream
>> commit 4f258a46):
>>
>>      sd: Fix maximum I/O size for BLOCK_PC requests
>>
>> With /etc/sysctl.conf containing
>>
>>      vm.dirty_background_bytes = 67108864
>>      vm.dirty_bytes = 1073741824
>>
>> a simple "dd" example writing 10GB file
>>
>>      dd if=/dev/zero of=ssd.test.file bs=1M count=10240
>>
>> results in about this on 3.18.21:
>>
>>      Dirty:            740856 kB
>>      Writeback:         12400 kB
>>
>> but on 3.18.22:
>>
>>      Dirty:             49244 kB
>>      Writeback:        656396 kB
>>
>> I.e. it seems to revert the relationship. I haven't identified any
>> performance impact, and apparently for random writes the behavior did
>> not change at all (or at least I haven't managed to reproduce it).
>>
>> But it's unclear to me why setting a maximum I/O size should affect
>> this, and perhaps it has impact that I don't see.
>
> So what appears to be happening here is that background writeback is
> cutting in earlier - the amount of pending writeback ("Dirty") is
> reduced while the amount of active writeback ("Writeback") is
> correspondingly increased.
>
> 4f258a46 had the effect of permitting larger requests into the
> request queue. It's unclear to me why larger requests would cause
> background writeback to cut in earlier - the writeback code doesn't
> even care about individual request sizes, it only cares about
> aggregate pagecache state.
>

Right. Not a kernel expert here, but that's mostly my thinking.

> Less Dirty and more Writeback isn't necessarily a bad thing at all,
> but I don't like mysteries. cc linux-mm to see if anyone else can
> spot-the-difference.
>

I'm not sure if the change has positive or negative impact (or perhaps 
no actual impact), but as a database guy (PostgreSQL) I'm interested in 
this, as the interaction between the database write activity and kernel 
matters to us a lot. So I'm wondering if this change might trigger the 
writeback sooner, etc.

regards
Tomas