Disappointing RAID10 Performance

Disappointing RAID10 Performance

[adfas asd]

I was hoping to get better performance with RAID10 than from the raw disks, but that's turned out to not be the case.  Experimenting with the readahead buffer I get these bandwidths with the following command:
# time dd if={somelarge}.iso of=/dev/null bs={readahead size}

/dev/sd?
 1024 71.3 MB/s
 2048 71.2 MB/s
 4096 77.7 MB/s
 8192 69.4 MB/s
 16384 76.6 MB/s

/dev/md2
 1024  67.1
 2048  69.1
 4096  75.7
 8192  64.9
 16384 69.0

Using RAID10offset2 on 2 WD 2TB drives, and always the same input file.

Why would RAID10 performance be -poorer-? 
If it weren't for mirroring, this wouldn't be worth it.


 


      

Re: Disappointing RAID10 Performance

[Majed B.]

Have you tried the tips in this page?
http://linux-raid.osdl.org/index.php/Performance#Some_problem_solving_for_benchmarking

On Fri, Oct 16, 2009 at 8:32 PM, adfas asd <chimera_god@yahoo.com> wrote:
> I was hoping to get better performance with RAID10 than from the raw disks, but that's turned out to not be the case.  Experimenting with the readahead buffer I get these bandwidths with the following command:
> # time dd if={somelarge}.iso of=/dev/null bs={readahead size}
>
> /dev/sd?
>  1024 71.3 MB/s
>  2048 71.2 MB/s
>  4096 77.7 MB/s
>  8192 69.4 MB/s
>  16384 76.6 MB/s
>
> /dev/md2
>  1024  67.1
>  2048  69.1
>  4096  75.7
>  8192  64.9
>  16384 69.0
>
> Using RAID10offset2 on 2 WD 2TB drives, and always the same input file.
>
> Why would RAID10 performance be -poorer-?
> If it weren't for mirroring, this wouldn't be worth it.
>
>
>
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>



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

Re: Disappointing RAID10 Performance

[Rob Becker]

    Hi,

   What command did you use to create your raid-10?  You might try  
running iostat in parallel to see if the read_balancer is properly  
balancing the reads between the two disks.

   I've seen situations where raid-10 on sequential reads will pretty  
much always prefer the same disk and barely use the mirror.

   One thing we looked at was changing to the far layout, which helped  
a bit.

--Rob
On Oct 16, 2009, at 10:32 AM, adfas asd wrote:

> I was hoping to get better performance with RAID10 than from the raw  
> disks, but that's turned out to not be the case.  Experimenting with  
> the readahead buffer I get these bandwidths with the following  
> command:
> # time dd if={somelarge}.iso of=/dev/null bs={readahead size}
>
> /dev/sd?
> 1024 71.3 MB/s
> 2048 71.2 MB/s
> 4096 77.7 MB/s
> 8192 69.4 MB/s
> 16384 76.6 MB/s
>
> /dev/md2
> 1024  67.1
> 2048  69.1
> 4096  75.7
> 8192  64.9
> 16384 69.0
>
> Using RAID10offset2 on 2 WD 2TB drives, and always the same input  
> file.
>
> Why would RAID10 performance be -poorer-?
> If it weren't for mirroring, this wouldn't be worth it.
>
>
>
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux- 
> raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: Disappointing RAID10 Performance

[adfas asd]

--- On Fri, 10/16/09, Rob Becker <Rob.Becker@riverbed.com> wrote:
>    What command did you use to create your
> raid-10?

/
mdadm --create /dev/md0 --level=raid1 --chunk=256 --raid-disks=2 missing /dev/sdb1
swap
mdadm --create /dev/md1 --level=raid10 --layout=o2 --chunk=256 --raid-disks=2 missing /dev/sdb2
/home
mdadm --create /dev/md2 --level=raid10 --layout=o2 --chunk=1024 --raid-disks=2 missing /dev/sdb3
... then copied files and later added the sda parts to the array. (RAID conversion on live system)


>  You might try  
> running iostat in parallel to see if the read_balancer is
> properly  
> balancing the reads between the two disks.

Don't understand this as I'm a bit of a n00b...


--- On Fri, 10/16/09, Tomasz Chmielewski <mangoo@wpkg.org> wrote:
> If you only use your RAID-10 array for a single "dd
> if=bigfile of=/dev/null" then yes, it does not give you much
> over mirroring.
> 
> If you start using your drives for two "dd if=bigfile[12]
> of=/dev/null" at the same time, you will notice the
> difference.

OK so it was a fallacy to think this would help with large files, unless more than one is involved.




      

Re: Disappointing RAID10 Performance

[Bill Davidsen]

adfas asd wrote:
> --- On Fri, 10/16/09, Rob Becker <Rob.Becker@riverbed.com> wrote:
>   
>>    What command did you use to create your
>> raid-10?
>>     
>
> /
> mdadm --create /dev/md0 --level=raid1 --chunk=256 --raid-disks=2 missing /dev/sdb1
> swap
> mdadm --create /dev/md1 --level=raid10 --layout=o2 --chunk=256 --raid-disks=2 missing /dev/sdb2
> /home
> mdadm --create /dev/md2 --level=raid10 --layout=o2 --chunk=1024 --raid-disks=2 missing /dev/sdb3
> ... then copied files and later added the sda parts to the array. (RAID conversion on live system)
>
>   
No wonder it's slow, you want two far copies, this is more or less 
mirroring with only two drives. Using a large buffer size also helps, 
you hurt your performance by limiting readahead. You can also use the 
'blockdev' command (--setra) to increase your readahead on the array. 
Just going to far should about double your speed, the other things may 
help more.
>   
>>  You might try  
>> running iostat in parallel to see if the read_balancer is
>> properly  
>> balancing the reads between the two disks.
>>     
>
> Don't understand this as I'm a bit of a n00b...
>
>
> --- On Fri, 10/16/09, Tomasz Chmielewski <mangoo@wpkg.org> wrote:
>   
>> If you only use your RAID-10 array for a single "dd
>> if=bigfile of=/dev/null" then yes, it does not give you much
>> over mirroring.
>>
>> If you start using your drives for two "dd if=bigfile[12]
>> of=/dev/null" at the same time, you will notice the
>> difference.
>>     
>
> OK so it was a fallacy to think this would help with large files, unless more than one is involved.
>   

You are misconfigured.

-- 
Bill Davidsen <davidsen@tmr.com>
  Unintended results are the well-earned reward for incompetence.

Re: Disappointing RAID10 Performance

[Tomasz Chmielewski]

adfas asd wrote:
> I was hoping to get better performance with RAID10 than from the raw disks, but that's turned out to not be the case.  Experimenting with the readahead buffer I get these bandwidths with the following command:
> # time dd if={somelarge}.iso of=/dev/null bs={readahead size}
> 
> /dev/sd?
>  1024 71.3 MB/s
>  2048 71.2 MB/s
>  4096 77.7 MB/s
>  8192 69.4 MB/s
>  16384 76.6 MB/s
> 
> /dev/md2
>  1024  67.1
>  2048  69.1
>  4096  75.7
>  8192  64.9
>  16384 69.0
> 
> Using RAID10offset2 on 2 WD 2TB drives, and always the same input file.
> 
> Why would RAID10 performance be -poorer-? 

If you only use your RAID-10 array for a single "dd if=bigfile 
of=/dev/null" then yes, it does not give you much over mirroring.

If you start using your drives for two "dd if=bigfile[12] of=/dev/null" 
at the same time, you will notice the difference.



-- 
Tomasz Chmielewski
http://wpkg.org

Re: Disappointing RAID10 Performance

[Asdo]

Tomasz Chmielewski wrote:
> If you start using your drives for two "dd if=bigfile[12] 
> of=/dev/null" at the same time, you will notice the difference.

OT: Does this dd line really work? I did see it elsewhere already but 
doesn't appear to work for me: seems to see only the last if=. I don't 
think multiple input or output files are supported.
For this kind of tests I was doing a for loop with background dd comands 
( trailing & ) IIRC
This would be a good dd feature to simplify linux-raid performance 
testing though.

Re: Disappointing RAID10 Performance

[Tomasz Chmielewski]

Asdo wrote:
> Tomasz Chmielewski wrote:
>> If you start using your drives for two "dd if=bigfile[12] 
>> of=/dev/null" at the same time, you will notice the difference.
> 
> OT: Does this dd line really work? I did see it elsewhere already but 
> doesn't appear to work for me: seems to see only the last if=. 

I don't think it does work.

You see it everywhere because it's faster to write, and everyone gets 
the idea what's this all about.


-- 
Tomasz Chmielewski
http://wpkg.org

Re: Disappointing RAID10 Performance

[Majed B.]

dd supports writing to multiple devices at once, but not reading from
multiple inputs.

On Sat, Oct 17, 2009 at 7:12 PM, Asdo <asdo@shiftmail.org> wrote:
> Tomasz Chmielewski wrote:
>>
>> If you start using your drives for two "dd if=bigfile[12] of=/dev/null" at
>> the same time, you will notice the difference.
>
> OT: Does this dd line really work? I did see it elsewhere already but
> doesn't appear to work for me: seems to see only the last if=. I don't think
> multiple input or output files are supported.
> For this kind of tests I was doing a for loop with background dd comands (
> trailing & ) IIRC
> This would be a good dd feature to simplify linux-raid performance testing
> though.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>



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

Re: Disappointing RAID10 Performance

[Michael Tokarev]

Majed B. Wrote:
> dd supports writing to multiple devices at once, but not reading from
> multiple inputs.

Count me intrigued.  I re-read linux dd manual page, POSIX docs about
dd and linux (coreutils) dd sources, and don't see the `multiple output'
thing.  Care to show how to do that?

Thanks.

/mjt

Re: Disappointing RAID10 Performance

[Christopher Chen]

Sorry, I've looked at this again.

It's plausible that md's read balancing would increase the read speed,
but it appears that this is not the case. Are you using LVM? And the
readahead for the block devices are set appropriately? Distro? Version
of mdadm?

cc

On Fri, Oct 16, 2009 at 10:32 AM, adfas asd <chimera_god@yahoo.com> wrote:
> I was hoping to get better performance with RAID10 than from the raw disks, but that's turned out to not be the case.  Experimenting with the readahead buffer I get these bandwidths with the following command:
> # time dd if={somelarge}.iso of=/dev/null bs={readahead size}
>
> /dev/sd?
>  1024 71.3 MB/s
>  2048 71.2 MB/s
>  4096 77.7 MB/s
>  8192 69.4 MB/s
>  16384 76.6 MB/s
>
> /dev/md2
>  1024  67.1
>  2048  69.1
>  4096  75.7
>  8192  64.9
>  16384 69.0
>
> Using RAID10offset2 on 2 WD 2TB drives, and always the same input file.
>
> Why would RAID10 performance be -poorer-?
> If it weren't for mirroring, this wouldn't be worth it.
>
>
>
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>



-- 
Chris Chen <muffaleta@gmail.com>
"The fact that yours is better than anyone else's
is not a guarantee that it's any good."
-- Seen on a wall
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: Disappointing RAID10 Performance

[adfas asd]

--- On Fri, 10/16/09, Christopher Chen <muffaleta@gmail.com> wrote:
> It's plausible that md's read balancing would increase the
> read speed,
> but it appears that this is not the case. Are you using
> LVM? And the
> readahead for the block devices are set appropriately?
> Distro? Version
> of mdadm?

No LVM, and readahead on sd? drives set for best performance.  Debian Testing with mdadm 3.0-2.


      

Re: Disappointing RAID10 Performance

[Keld Jørn Simonsen]

On Fri, Oct 16, 2009 at 10:32:14AM -0700, adfas asd wrote:
> I was hoping to get better performance with RAID10 than from the raw disks, but that's turned out to not be the case.  Experimenting with the readahead buffer I get these bandwidths with the following command:
> # time dd if={somelarge}.iso of=/dev/null bs={readahead size}
> 
> /dev/sd?
>  1024 71.3 MB/s
>  2048 71.2 MB/s
>  4096 77.7 MB/s
>  8192 69.4 MB/s
>  16384 76.6 MB/s
> 
> /dev/md2
>  1024  67.1
>  2048  69.1
>  4096  75.7
>  8192  64.9
>  16384 69.0
> 
> Using RAID10offset2 on 2 WD 2TB drives, and always the same input file.
> 
> Why would RAID10 performance be -poorer-? 
> If it weren't for mirroring, this wouldn't be worth it.

try with layout=far - it is good for reading big files, maybe a factor 2
faster than layout=offset.

More on performance:
http://linux-raid.osdl.org/index.php/Performance

Best regards
keld

RE: Disappointing RAID10 Performance

[Leslie Rhorer]

> # time dd if={somelarge}.iso of=/dev/null bs={readahead size}
> 
> /dev/sd?
>  1024 71.3 MB/s
>  2048 71.2 MB/s
>  4096 77.7 MB/s
>  8192 69.4 MB/s
>  16384 76.6 MB/s
> 
> /dev/md2
>  1024  67.1
>  2048  69.1
>  4096  75.7
>  8192  64.9
>  16384 69.0
> 
> Using RAID10offset2 on 2 WD 2TB drives, and always the same input file.

	RAID10 on 2 drives?  Are you striping partitions on a single drive?
You won't get much of a performance boost that way.

> Why would RAID10 performance be -poorer-?
> If it weren't for mirroring, this wouldn't be worth it.

Is this with all drives local, or are you mirroring across a Gig-E link, as
I believe you said you were going to do?

RE: Disappointing RAID10 Performance

[adfas asd]

--- On Fri, 10/16/09, Leslie Rhorer <lrhorer@satx.rr.com> wrote:
> > Why would RAID10 performance be -poorer-?
> > If it weren't for mirroring, this wouldn't be worth
> it.
> 
> Is this with all drives local, or are you mirroring across
> a Gig-E link, as
> I believe you said you were going to do?

Do not have the remote array running yet, so all drives are local.



      

Re: Disappointing RAID10 Performance

[adfas asd]

> > Yeah two drives.  I didn't know I could have a RAID10
> > with 3 drives?  I thought I always had to add 2?  I'm
> > afraid to ask how it would set up an offset2 layout with
> > 3...
> Well you have to put it up on a
> whiteboard. It means that every data
> block exists as a "mirror" on any two drives. Hrm. It's
> fun. Draw it
> out :)

But to add a drive I have to tear down the array and rebuild it, right?  This gets impossible with the amount of data in larger arrays.

And I guess this isn't really relevant with my system, where one side of the mirror is in the HTPC, and the other is in the garage.  A fire in either place would destroy all data in an odd number drive system.