Performance comparaison with dbench on KVM

Performance comparaison with dbench on KVM

[Laurent Vivier]

[-- Attachment #1.1: Type: text/plain, Size: 1120 bytes --]

Hi,

I made some benchmarks with dbench on linux/KVM.

IDE means I use an image disk with "-hda"
SCSI means I use an image disk with "-sda"
(see the patch I proposed on qemu list:
http://www.mail-archive.com/qemu-devel-qX2TKyscuCcdnm+yROfE0A@public.gmane.org/msg08679.html )

threads KVM-21  KVM-21  KVM-24  KVM-24  KVM-25  KVM-25
        IDE     SCSI    IDE     SCSI    IDE     SCSI

1       215.272 94.8775 318.626 319.533 307.111 331.757
2       212.061 91.3193 313.217 315.163 314.577 323.715
4       197.586 81.7551 291.752 276.044 292.084 271.191
8       144.814 58.5725 188.819 198.995 168.273 165.564
16      111.505 66.3605 139.824 157.617 150.822 123.671
32      43.4109 15.7386 48.2109 50.4114 59.1074 69.5717
64      21.0567 8.3535  11.3457 19.9483 20.4877 24.5627

	results are in MB/sec

Conclusion:

KVM-25 is really better than KVM-21
in most cases, SCSI is better than IDE

Cheers,
Laurent
-- 
------------- Laurent.Vivier-6ktuUTfB/bM@public.gmane.org  --------------
       "Any sufficiently advanced technology is
  indistinguishable from magic." - Arthur C. Clarke


[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

[-- Attachment #2: Type: text/plain, Size: 286 bytes --]

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

[-- Attachment #3: Type: text/plain, Size: 186 bytes --]

_______________________________________________
kvm-devel mailing list
kvm-devel-5NWGOfrQmneRv+LV9MX5uipxlwaOVQ5f@public.gmane.org
https://lists.sourceforge.net/lists/listinfo/kvm-devel

Re: Performance comparaison with dbench on KVM

[Anthony Liguori]

Laurent Vivier wrote:
> Hi,
>
> I made some benchmarks with dbench on linux/KVM.
>
> IDE means I use an image disk with "-hda"
> SCSI means I use an image disk with "-sda"
> (see the patch I proposed on qemu list:
> http://www.mail-archive.com/qemu-devel-qX2TKyscuCcdnm+yROfE0A@public.gmane.org/msg08679.html )
>
> threads KVM-21  KVM-21  KVM-24  KVM-24  KVM-25  KVM-25
>         IDE     SCSI    IDE     SCSI    IDE     SCSI
>
> 1       215.272 94.8775 318.626 319.533 307.111 331.757
> 2       212.061 91.3193 313.217 315.163 314.577 323.715
> 4       197.586 81.7551 291.752 276.044 292.084 271.191
> 8       144.814 58.5725 188.819 198.995 168.273 165.564
> 16      111.505 66.3605 139.824 157.617 150.822 123.671
> 32      43.4109 15.7386 48.2109 50.4114 59.1074 69.5717
> 64      21.0567 8.3535  11.3457 19.9483 20.4877 24.5627
>
> 	results are in MB/sec
>
> Conclusion:
>
> KVM-25 is really better than KVM-21
> in most cases, SCSI is better than IDE
>   

Were you using libposix-aio?  If not, did you change the aio_init 
function in block-raw.c (assuming that you're using a raw disk).  
Otherwise, QEMU by default will limit the glibc aio library to a single 
thread so you won't see a lot of the benefits of SCSI emulation.

Regards,

Anthony Liguori

> Cheers,
> Laurent
>   
> ------------------------------------------------------------------------
>
> -------------------------------------------------------------------------
> This SF.net email is sponsored by DB2 Express
> Download DB2 Express C - the FREE version of DB2 express and take
> control of your XML. No limits. Just data. Click to get it now.
> http://sourceforge.net/powerbar/db2/
> ------------------------------------------------------------------------
>
> _______________________________________________
> kvm-devel mailing list
> kvm-devel-5NWGOfrQmneRv+LV9MX5uipxlwaOVQ5f@public.gmane.org
> https://lists.sourceforge.net/lists/listinfo/kvm-devel
>   


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: Performance comparaison with dbench on KVM

[Laurent Vivier]

[-- Attachment #1.1: Type: text/plain, Size: 1859 bytes --]

Anthony Liguori wrote:
> Laurent Vivier wrote:
>> Hi,
>>
>> I made some benchmarks with dbench on linux/KVM.
>>
>> IDE means I use an image disk with "-hda"
>> SCSI means I use an image disk with "-sda"
>> (see the patch I proposed on qemu list:
>> http://www.mail-archive.com/qemu-devel-qX2TKyscuCcdnm+yROfE0A@public.gmane.org/msg08679.html )
>>
>> threads KVM-21  KVM-21  KVM-24  KVM-24  KVM-25  KVM-25
>>         IDE     SCSI    IDE     SCSI    IDE     SCSI
>>
>> 1       215.272 94.8775 318.626 319.533 307.111 331.757
>> 2       212.061 91.3193 313.217 315.163 314.577 323.715
>> 4       197.586 81.7551 291.752 276.044 292.084 271.191
>> 8       144.814 58.5725 188.819 198.995 168.273 165.564
>> 16      111.505 66.3605 139.824 157.617 150.822 123.671
>> 32      43.4109 15.7386 48.2109 50.4114 59.1074 69.5717
>> 64      21.0567 8.3535  11.3457 19.9483 20.4877 24.5627
>>
>> 	results are in MB/sec
>>
>> Conclusion:
>>
>> KVM-25 is really better than KVM-21
>> in most cases, SCSI is better than IDE
>>   
> 
> Were you using libposix-aio?  If not, did you change the aio_init 
> function in block-raw.c (assuming that you're using a raw disk).  
> Otherwise, QEMU by default will limit the glibc aio library to a single 
> thread so you won't see a lot of the benefits of SCSI emulation.

Hi Anthony,

I didn't modify aio_init limit...

I made some benchmarks with liposix-aio, but results are worst with IDE and
crashes with SCSI (because I need to make some additional works).

It's the part on which I work now (to align buffers to avoid redundant buffer
copy and open image disk with O_DIRECT).

Regards,
Laurent

-- 
------------- Laurent.Vivier-6ktuUTfB/bM@public.gmane.org  --------------
       "Any sufficiently advanced technology is
  indistinguishable from magic." - Arthur C. Clarke


[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

[-- Attachment #2: Type: text/plain, Size: 286 bytes --]

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

[-- Attachment #3: Type: text/plain, Size: 186 bytes --]

_______________________________________________
kvm-devel mailing list
kvm-devel-5NWGOfrQmneRv+LV9MX5uipxlwaOVQ5f@public.gmane.org
https://lists.sourceforge.net/lists/listinfo/kvm-devel

Re: Performance comparaison with dbench on KVM

[Anthony Liguori]

Laurent Vivier wrote:
> Anthony Liguori wrote:
>   
> Hi Anthony,
>
> I didn't modify aio_init limit...
>   

It would be pretty useful to give it a try.  In qemu_aio_init() where 
aio_init() get's called.  If you just bump aio_threads and aio_num from 
1 to, say, 32. You should see a big difference in SCSI performance.

Regards,

Anthony Liguori

> I made some benchmarks with liposix-aio, but results are worst with IDE and
> crashes with SCSI (because I need to make some additional works).
>
> It's the part on which I work now (to align buffers to avoid redundant buffer
> copy and open image disk with O_DIRECT).
>
> Regards,
> Laurent
>
>   


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: Performance comparaison with dbench on KVM

[Laurent Vivier]

[-- Attachment #1.1: Type: text/plain, Size: 992 bytes --]

Anthony Liguori wrote:
> Laurent Vivier wrote:
>> Anthony Liguori wrote:
>>   Hi Anthony,
>>
>> I didn't modify aio_init limit...
>>   
> 
> It would be pretty useful to give it a try.  In qemu_aio_init() where
> aio_init() get's called.  If you just bump aio_threads and aio_num from
> 1 to, say, 32. You should see a big difference in SCSI performance.
> 

I made some tests as you explained, but results are worst.

with:
        ai.aio_threads = 32;
        ai.aio_num = 32;

I have:

KVM-25 + SCSI + aio_init(32 threads)

Throughput 332.305 MB/sec 1 procs
Throughput 316.832 MB/sec 2 procs
Throughput 246.741 MB/sec 4 procs
Throughput 142.115 MB/sec 8 procs
Throughput 107.227 MB/sec 16 procs
Throughput 54.6614 MB/sec 32 proc
Throughput 21.3738 MB/sec 64 procs

Laurent
-- 
------------- Laurent.Vivier-6ktuUTfB/bM@public.gmane.org  --------------
       "Any sufficiently advanced technology is
  indistinguishable from magic." - Arthur C. Clarke


[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

[-- Attachment #2: Type: text/plain, Size: 286 bytes --]

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

[-- Attachment #3: Type: text/plain, Size: 186 bytes --]

_______________________________________________
kvm-devel mailing list
kvm-devel-5NWGOfrQmneRv+LV9MX5uipxlwaOVQ5f@public.gmane.org
https://lists.sourceforge.net/lists/listinfo/kvm-devel