Re: Btrfs performance with small blocksize on SSD

[Peter Grandi]

> i run a few performance tests comparing mdadm, hardware raid
> and the btrfs raid.

Fantastic beginning already! :-)

> I noticed that the performance

I have seen over the years a lot of messages like this where
there is a wanton display of amusing misuses of terminology, of
which the misuse of the word "performance" to mean "speed" is
common, and your results are work-per-time which is a "speed":
http://www.sabi.co.uk/blog/15-two.html?151023#151023

The "tl;dr" is: you and another guy are told to race the 100m to
win a €10,000 prize, but you have to carry a sack with a 50Kg
weight. It takes you a lot longer, as your speed is much lower,
and the other guy gets the prize. Was that because your
performance was much worse? :-)

> for small blocksizes (2k) is very bad on SSD in general and on
> HDD for sequential writing.

Your graphs show pretty decent performance for small-file IO on
Btrfs, depending on conditions, and you are very astutely not
explaining the conditions, even if some can be guessed.

> I wonder about that result, because you say on the wiki that
> btrfs is very effective for small files.

Effectivess/efficiency are not the same as performance or speed
either. My own simplistic but somewhat meaningful tests show
that Btrfs does relatively well on small files:

  http://www.sabi.co.uk/blog/17-one.html?170302#170302

As to "small files" in general I have read about many attempts
to use filesystems as DBMSes, and I consider them intensely
stupid:

  http://www.sabi.co.uk/blog/anno05-4th.html?051016#051016

> I attached my results from raid 1 random write HDD (rH1), SSD
> (rS1) and from sequential write HDD (sH1), SSD (sS1)

Ah, so it was specifically about small *writes* (and presumably
because of other wording not small-updates-in-place of large
files, but creating and writing small files).

It is a very basic beginner level notion that most storage
systems are very anisotropic as to IO size, and also for read
vs. write, and never mind with and without 'fsync'. SSDs without
supercapacitor backed buffers in particular are an issue.

Btrfs has a performance envelope where the speed of small writes
(in particular small in-place updates, but also because of POSIX
small file creation) has been sacrificed for good reasons:

https://btrfs.wiki.kernel.org/index.php/SysadminGuide#Copy_on_Write_.28CoW.29
https://btrfs.wiki.kernel.org/index.php/Gotchas#Fragmentation

Also consider the consequences of the 'max_inline' option for
'mount' and the 'nodesize' option for 'mkfs.btrfs'.

> Hopefully you have an explanation for that.

The best explanation seems to me (euphemism alert) quite
extensive "misknowledge" in the message I am responding to.
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: Btrfs performance with small blocksize on SSD

[Kai Krakow]

Am Mon, 25 Sep 2017 07:04:14 +
schrieb "Fuhrmann, Carsten" <carsten.fuhrm...@rwth-aachen.de>:

> Well the correct translation for "Laufzeit" is runtime and not
> latency. But thank you for that hint, I'll change it to "gesamt
> Laufzeit" to make it more clear.

How about better translating it to English in the first place as you're
trying to reach an international community?

Also, it would be nice to put the exact test you did as a command line
or configuration file, so it can be replayed on other systems, and -
most important - by the developers, to easily uncover what is causing
the behavior...


> Best regards
> 
> Carsten
> 
> -Ursprüngliche Nachricht-
> Von: Andrei Borzenkov [mailto:arvidj...@gmail.com] 
> Gesendet: Sonntag, 24. September 2017 18:43
> An: Fuhrmann, Carsten <carsten.fuhrm...@rwth-aachen.de>; Qu Wenruo
> <quwenruo.bt...@gmx.com>; linux-btrfs@vger.kernel.org Betreff: Re:
> AW: Btrfs performance with small blocksize on SSD
> 
> 24.09.2017 16:53, Fuhrmann, Carsten пишет:
> > Hello,
> > 
> > 1)
> > I used direct write (no page cache) but I didn't disable the Disk
> > cache of the HDD/SSD itself. In all tests I wrote 1GB and looked
> > for the runtime of that write process.  
> 
> So "latency" on your diagram means total time to write 1GiB file?
> That is highly unusual meaning for "latency" which normally means
> time to perform single IO. If so, you should better rename Y-axis to
> something like "total run time".
> N_r__yb_Xv_^_)__{.n_+{_n___)w*jg______j/___z_2___&_)___a_____G___h__j:+v___w


-- 
Regards,
Kai

Replies to list-only preferred.


--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

AW: AW: Btrfs performance with small blocksize on SSD

[Fuhrmann, Carsten]

Well the correct translation for "Laufzeit" is runtime and not latency. But 
thank you for that hint, I'll change it to "gesamt Laufzeit" to make it more 
clear.


Best regards

Carsten

-Ursprüngliche Nachricht-
Von: Andrei Borzenkov [mailto:arvidj...@gmail.com] 
Gesendet: Sonntag, 24. September 2017 18:43
An: Fuhrmann, Carsten <carsten.fuhrm...@rwth-aachen.de>; Qu Wenruo 
<quwenruo.bt...@gmx.com>; linux-btrfs@vger.kernel.org
Betreff: Re: AW: Btrfs performance with small blocksize on SSD

24.09.2017 16:53, Fuhrmann, Carsten пишет:
> Hello,
> 
> 1)
> I used direct write (no page cache) but I didn't disable the Disk cache of 
> the HDD/SSD itself. In all tests I wrote 1GB and looked for the runtime of 
> that write process.

So "latency" on your diagram means total time to write 1GiB file? That is 
highly unusual meaning for "latency" which normally means time to perform 
single IO. If so, you should better rename Y-axis to something like "total run 
time".
N�r��yb�X��ǧv�^�)޺{.n�+{�n�߲)w*jg����ݢj/���z�ޖ��2�ޙ&�)ߡ�a�����G���h��j:+v���w��٥

Re: Btrfs performance with small blocksize on SSD

[Kai Krakow]

Am Sun, 24 Sep 2017 19:43:05 +0300
schrieb Andrei Borzenkov :

> 24.09.2017 16:53, Fuhrmann, Carsten пишет:
> > Hello,
> > 
> > 1)
> > I used direct write (no page cache) but I didn't disable the Disk
> > cache of the HDD/SSD itself. In all tests I wrote 1GB and looked
> > for the runtime of that write process.  
> 
> So "latency" on your diagram means total time to write 1GiB file? That
> is highly unusual meaning for "latency" which normally means time to
> perform single IO. If so, you should better rename Y-axis to something
> like "total run time".

If you look closely it says "Laufzeit" which visually looks similar to
"latency" but really means "run time". ;-)


-- 
Regards,
Kai

Replies to list-only preferred.


--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: AW: Btrfs performance with small blocksize on SSD

[Andrei Borzenkov]

24.09.2017 16:53, Fuhrmann, Carsten пишет:
> Hello,
> 
> 1)
> I used direct write (no page cache) but I didn't disable the Disk cache of 
> the HDD/SSD itself. In all tests I wrote 1GB and looked for the runtime of 
> that write process.

So "latency" on your diagram means total time to write 1GiB file? That
is highly unusual meaning for "latency" which normally means time to
perform single IO. If so, you should better rename Y-axis to something
like "total run time".
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

AW: AW: Btrfs performance with small blocksize on SSD

[Fuhrmann, Carsten]

1)
Every test has it's own file. So the 2k blocksize write to a different file 
then the 4k blocksize test. In the End there are 5 files on the disk (2k, 
8k,...)

2)
Well I think it is 2 as well since for 4k and higher the performance is much 
better .
I'm gonna test the -o max_inline and test with 4k blocksize as well.

Thank you for your help

Carsten

-Ursprüngliche Nachricht-
Von: Qu Wenruo [mailto:quwenruo.bt...@gmx.com] 
Gesendet: Sonntag, 24. September 2017 16:11
An: Fuhrmann, Carsten <carsten.fuhrm...@rwth-aachen.de>; 
linux-btrfs@vger.kernel.org
Betreff: Re: AW: Btrfs performance with small blocksize on SSD



On 2017年09月24日 21:53, Fuhrmann, Carsten wrote:
> Hello,
> 
> 1)
> I used direct write (no page cache) but I didn't disable the Disk cache of 
> the HDD/SSD itself. In all tests I wrote 1GB and looked for the runtime of 
> that write process.

Are you writing all the 1G into one file?
Or into different files?

> I run every test 5 times with different Blocksizes (2k, 8k, 32k, 128k, 512k). 
> Those values are on the x-axis. On the Y-Axis is the runtime for the test.

Good to know that.

Then there may be 2 factors impacting performance:

1) Convert between inlined and regular data extent
1st 2K write will be inlined and then 2nd 2K write will convert it
back to regular data extent.
The overhead can be quite high.

Retest with "-o max_inline=0" will disable such behavior so all write
will only cause regular data extent.

2) Unaligned data size causing extra rewrite/CoW
Btrfs restore its data in unit of sectorsize, and in your case it is
4K.
Writing with 2K will cause btrfs to read out the half written data
and then CoW it to somewhere else.
The overhead can be quite huge.

And I assume 2) is the main overhead.

Retest with 4K blocksize to see if it's related.

Please note that, 4K blocksize and 2K blocksize are going through
different routines (4K blocks routine has no extra CoW overhead, so
it' should be near 8K blockszie result)

Thanks,
Qu

> 
> 2)
> Yes every test is on RAID1 for data and metadata
> 
> 3)
> Everything default
> mkfs.btrfs -d raid1 -m raid1 /dev/sda /dev/sdb /dev/sdc /dev/sdd
> 
> 
> best regards
> 
> Carsten
> 
> -Ursprüngliche Nachricht-
> Von: Qu Wenruo [mailto:quwenruo.bt...@gmx.com]
> Gesendet: Sonntag, 24. September 2017 15:41
> An: Fuhrmann, Carsten <carsten.fuhrm...@rwth-aachen.de>; 
> linux-btrfs@vger.kernel.org
> Betreff: Re: Btrfs performance with small blocksize on SSD
> 
> 
> 
> On 2017年09月24日 21:24, Fuhrmann, Carsten wrote:
>> Hello,
>>
>> i run a few performance tests comparing mdadm, hardware raid and the btrfs 
>> raid. I noticed that the performance for small blocksizes (2k) is very bad 
>> on SSD in general and on HDD for sequential writing.
> 
> 2K is smaller than the minimal btrfs sectorsize (4K for x86 family).
> 
> It's common that unaligned access will impact performance, but we need more 
> info about your test cases, including:
> 1) How write is done?
>  Buffered? DIO? O_SYNC? fdatasync?
>  I can't read Germany so I'm not sure what the result means. (Although
>  I can guess Y axle is latency, but I don't know the meaning of X axle.
>  And how many files are involved, how large of these files and etc.
> 
> 2) Data/meta/sys profiles
>  All RADI1?
> 
> 3) Mkfs profile
>  Like nodesize if not default, and any incompat features enabled.
> 
>> I wonder about that result, because you say on the wiki that btrfs is very 
>> effective for small files.
> 
> It can be space effective or performance effective.
> 
> If *ignoring* meta profile, btrfs is space-effectient since it inline the 
> data into metadata, avoiding padding it to sectorsize so it can save some 
> space.
> 
> And such behavior can also be somewhat performance effective, by avoiding 
> extra seeking for data, since when reading out the metadata we have already 
> read out the inlined data.
> 
> But such efficiency come with cost.
> 
> One obvious one is when we need to convert inline data into regular one.
> It may cause extra tree balancing and increase latency.
> 
> Would you please try retest with "-o max_inline=0" mount option to disable 
> inline data (which makes btrfs behavior like ext*/xfs) to see if it's related?
> 
> Thanks,
> Qu
> 
>>
>> I attached my results from raid 1 random write HDD (rH1), SSD (rS1) 
>> and from sequential write HDD (sH1), SSD (sS1)
>>
>> Hopefully you have an explanation for that.
>>
>> raid@raid-PowerEdge-T630:~$ uname -a
>> Linux raid-PowerEdge-T630 4.10.0-33-generic #37~16.04.1-Ubuntu SMP 
>> Fri Aug 11 14:07:24 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux 
>> raid@raid-PowerEdge-T630:~$ btrfs --version btrfs-progs v4.4
>>
>>
>> best regards
>>
>> Carsten
>>
> N r  y   b X  ǧv ^ )޺{.n +{ n ߲)   w*jg    ݢj/   z ޖ  2 
> ޙ   & )ߡ a     G   h  j:+v   w ٥
> 

Re: AW: Btrfs performance with small blocksize on SSD

[Qu Wenruo]

On 2017年09月24日 21:53, Fuhrmann, Carsten wrote:

Hello,

1)
I used direct write (no page cache) but I didn't disable the Disk cache of the 
HDD/SSD itself. In all tests I wrote 1GB and looked for the runtime of that 
write process.


Are you writing all the 1G into one file?
Or into different files?


I run every test 5 times with different Blocksizes (2k, 8k, 32k, 128k, 512k). 
Those values are on the x-axis. On the Y-Axis is the runtime for the test.


Good to know that.

Then there may be 2 factors impacting performance:

1) Convert between inlined and regular data extent
   1st 2K write will be inlined and then 2nd 2K write will convert it
   back to regular data extent.
   The overhead can be quite high.

   Retest with "-o max_inline=0" will disable such behavior so all write
   will only cause regular data extent.

2) Unaligned data size causing extra rewrite/CoW
   Btrfs restore its data in unit of sectorsize, and in your case it is
   4K.
   Writing with 2K will cause btrfs to read out the half written data
   and then CoW it to somewhere else.
   The overhead can be quite huge.

   And I assume 2) is the main overhead.

   Retest with 4K blocksize to see if it's related.

   Please note that, 4K blocksize and 2K blocksize are going through
   different routines (4K blocks routine has no extra CoW overhead, so
   it' should be near 8K blockszie result)

Thanks,
Qu



2)
Yes every test is on RAID1 for data and metadata

3)
Everything default
mkfs.btrfs -d raid1 -m raid1 /dev/sda /dev/sdb /dev/sdc /dev/sdd


best regards

Carsten

-Ursprüngliche Nachricht-
Von: Qu Wenruo [mailto:quwenruo.bt...@gmx.com]
Gesendet: Sonntag, 24. September 2017 15:41
An: Fuhrmann, Carsten <carsten.fuhrm...@rwth-aachen.de>; 
linux-btrfs@vger.kernel.org
Betreff: Re: Btrfs performance with small blocksize on SSD



On 2017年09月24日 21:24, Fuhrmann, Carsten wrote:

Hello,

i run a few performance tests comparing mdadm, hardware raid and the btrfs 
raid. I noticed that the performance for small blocksizes (2k) is very bad on 
SSD in general and on HDD for sequential writing.


2K is smaller than the minimal btrfs sectorsize (4K for x86 family).

It's common that unaligned access will impact performance, but we need more 
info about your test cases, including:
1) How write is done?
 Buffered? DIO? O_SYNC? fdatasync?
 I can't read Germany so I'm not sure what the result means. (Although
 I can guess Y axle is latency, but I don't know the meaning of X axle.
 And how many files are involved, how large of these files and etc.

2) Data/meta/sys profiles
 All RADI1?

3) Mkfs profile
 Like nodesize if not default, and any incompat features enabled.


I wonder about that result, because you say on the wiki that btrfs is very 
effective for small files.


It can be space effective or performance effective.

If *ignoring* meta profile, btrfs is space-effectient since it inline the data 
into metadata, avoiding padding it to sectorsize so it can save some space.

And such behavior can also be somewhat performance effective, by avoiding extra 
seeking for data, since when reading out the metadata we have already read out 
the inlined data.

But such efficiency come with cost.

One obvious one is when we need to convert inline data into regular one.
It may cause extra tree balancing and increase latency.

Would you please try retest with "-o max_inline=0" mount option to disable 
inline data (which makes btrfs behavior like ext*/xfs) to see if it's related?

Thanks,
Qu



I attached my results from raid 1 random write HDD (rH1), SSD (rS1)
and from sequential write HDD (sH1), SSD (sS1)

Hopefully you have an explanation for that.

raid@raid-PowerEdge-T630:~$ uname -a
Linux raid-PowerEdge-T630 4.10.0-33-generic #37~16.04.1-Ubuntu SMP Fri
Aug 11 14:07:24 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux
raid@raid-PowerEdge-T630:~$ btrfs --version btrfs-progs v4.4


best regards

Carsten


N�r��y���b�X��ǧv�^�)޺{.n�+{�n�߲)���w*jg����ݢj/���z�ޖ��2�ޙ���&�)ߡ�a�����G���h��j:+v���w�٥


--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

AW: Btrfs performance with small blocksize on SSD

[Fuhrmann, Carsten]

Hello,

1)
I used direct write (no page cache) but I didn't disable the Disk cache of the 
HDD/SSD itself. In all tests I wrote 1GB and looked for the runtime of that 
write process.
I run every test 5 times with different Blocksizes (2k, 8k, 32k, 128k, 512k). 
Those values are on the x-axis. On the Y-Axis is the runtime for the test.

2)
Yes every test is on RAID1 for data and metadata

3)
Everything default
mkfs.btrfs -d raid1 -m raid1 /dev/sda /dev/sdb /dev/sdc /dev/sdd


best regards

Carsten

-Ursprüngliche Nachricht-
Von: Qu Wenruo [mailto:quwenruo.bt...@gmx.com] 
Gesendet: Sonntag, 24. September 2017 15:41
An: Fuhrmann, Carsten <carsten.fuhrm...@rwth-aachen.de>; 
linux-btrfs@vger.kernel.org
Betreff: Re: Btrfs performance with small blocksize on SSD



On 2017年09月24日 21:24, Fuhrmann, Carsten wrote:
> Hello,
> 
> i run a few performance tests comparing mdadm, hardware raid and the btrfs 
> raid. I noticed that the performance for small blocksizes (2k) is very bad on 
> SSD in general and on HDD for sequential writing.

2K is smaller than the minimal btrfs sectorsize (4K for x86 family).

It's common that unaligned access will impact performance, but we need more 
info about your test cases, including:
1) How write is done?
Buffered? DIO? O_SYNC? fdatasync?
I can't read Germany so I'm not sure what the result means. (Although
I can guess Y axle is latency, but I don't know the meaning of X axle.
And how many files are involved, how large of these files and etc.

2) Data/meta/sys profiles
All RADI1?

3) Mkfs profile
Like nodesize if not default, and any incompat features enabled.

> I wonder about that result, because you say on the wiki that btrfs is very 
> effective for small files.

It can be space effective or performance effective.

If *ignoring* meta profile, btrfs is space-effectient since it inline the data 
into metadata, avoiding padding it to sectorsize so it can save some space.

And such behavior can also be somewhat performance effective, by avoiding extra 
seeking for data, since when reading out the metadata we have already read out 
the inlined data.

But such efficiency come with cost.

One obvious one is when we need to convert inline data into regular one.
It may cause extra tree balancing and increase latency.

Would you please try retest with "-o max_inline=0" mount option to disable 
inline data (which makes btrfs behavior like ext*/xfs) to see if it's related?

Thanks,
Qu

> 
> I attached my results from raid 1 random write HDD (rH1), SSD (rS1) 
> and from sequential write HDD (sH1), SSD (sS1)
> 
> Hopefully you have an explanation for that.
> 
> raid@raid-PowerEdge-T630:~$ uname -a
> Linux raid-PowerEdge-T630 4.10.0-33-generic #37~16.04.1-Ubuntu SMP Fri 
> Aug 11 14:07:24 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux 
> raid@raid-PowerEdge-T630:~$ btrfs --version btrfs-progs v4.4
> 
> 
> best regards
> 
> Carsten
> 

Re: Btrfs performance with small blocksize on SSD

[Qu Wenruo]

On 2017年09月24日 21:24, Fuhrmann, Carsten wrote:

Hello,

i run a few performance tests comparing mdadm, hardware raid and the btrfs 
raid. I noticed that the performance for small blocksizes (2k) is very bad on 
SSD in general and on HDD for sequential writing.


2K is smaller than the minimal btrfs sectorsize (4K for x86 family).

It's common that unaligned access will impact performance, but we need 
more info about your test cases, including:

1) How write is done?
   Buffered? DIO? O_SYNC? fdatasync?
   I can't read Germany so I'm not sure what the result means. (Although
   I can guess Y axle is latency, but I don't know the meaning of X axle.
   And how many files are involved, how large of these files and etc.

2) Data/meta/sys profiles
   All RADI1?

3) Mkfs profile
   Like nodesize if not default, and any incompat features enabled.


I wonder about that result, because you say on the wiki that btrfs is very 
effective for small files.


It can be space effective or performance effective.

If *ignoring* meta profile, btrfs is space-effectient since it inline 
the data into metadata, avoiding padding it to sectorsize so it can save 
some space.


And such behavior can also be somewhat performance effective, by 
avoiding extra seeking for data, since when reading out the metadata we 
have already read out the inlined data.


But such efficiency come with cost.

One obvious one is when we need to convert inline data into regular one.
It may cause extra tree balancing and increase latency.

Would you please try retest with "-o max_inline=0" mount option to 
disable inline data (which makes btrfs behavior like ext*/xfs) to see if 
it's related?


Thanks,
Qu



I attached my results from raid 1 random write HDD (rH1), SSD (rS1) and from 
sequential write HDD (sH1), SSD (sS1)

Hopefully you have an explanation for that.

raid@raid-PowerEdge-T630:~$ uname -a
Linux raid-PowerEdge-T630 4.10.0-33-generic #37~16.04.1-Ubuntu SMP Fri Aug 11 
14:07:24 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux
raid@raid-PowerEdge-T630:~$ btrfs --version
btrfs-progs v4.4


best regards

Carsten


--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Btrfs performance with small blocksize on SSD

[Fuhrmann, Carsten]

Hello,

i run a few performance tests comparing mdadm, hardware raid and the btrfs 
raid. I noticed that the performance for small blocksizes (2k) is very bad on 
SSD in general and on HDD for sequential writing.
I wonder about that result, because you say on the wiki that btrfs is very 
effective for small files. 

I attached my results from raid 1 random write HDD (rH1), SSD (rS1) and from 
sequential write HDD (sH1), SSD (sS1)

Hopefully you have an explanation for that.

raid@raid-PowerEdge-T630:~$ uname -a
Linux raid-PowerEdge-T630 4.10.0-33-generic #37~16.04.1-Ubuntu SMP Fri Aug 11 
14:07:24 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux
raid@raid-PowerEdge-T630:~$ btrfs --version
btrfs-progs v4.4


best regards

Carsten