performance measurement

[gr3go]

Hy all!

Did anyone tested equalizer on bigger clusters like 10+ nodes?
I don't have the possibility to benchmark on more nodes, so I would appreciate if anyone could provide some info related to this issue.

Thanks,
Adam

P.S.: Congrats for the Equalizer team for efforts, developing this nice toolkit!

[Maxim Makhinya]

Hi,


I did some testing for our EQ paper (1..14-16 nodes; 1 GPU per
node), but it was almost a year ago. You can find this paper here:

http://ieeexplore.ieee.org/search/wrapper.jsp?arnumber=4624254

Also I've just started testing EQ on another cluster with 10 nodes,
2-4 GPU per node, but can't provide you any feedback yet.


What is your main application for parallel rendering? What kind of
a setup you are interested in?


Best regards,

Makhinya Maxim



On May 6, 2009, at 2:18 PM, gr3go wrote:

>
> Hy all!
>
> Did anyone tested equalizer on bigger clusters like 10+ nodes?
> I don't have the possibility to benchmark on more nodes, so I would
> appreciate if anyone could provide some info related to this issue.
>
> Thanks,
> Adam
>
> P.S.: Congrats for the Equalizer team for efforts, developing this  
> nice
> toolkit!
> --
> View this message in context: http://n2.nabble.com/performance-measurement-tp2813299p2813299.html
> Sent from the Equalizer - Parallel Rendering mailing list archive at  
> Nabble.com.
>
>
> _______________________________________________
> eq-dev mailing list
> [hidden email]
> http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
> http://www.equalizergraphics.com
>

_______________________________________________
eq-dev mailing list
[hidden email]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

[gr3go]

Maxim Makhinya wrote:

Hi,


I did some testing for our EQ paper (1..14-16 nodes; 1 GPU per
node), but it was almost a year ago. You can find this paper here:

http://ieeexplore.ieee.org/search/wrapper.jsp?arnumber=4624254

Also I've just started testing EQ on another cluster with 10 nodes,
2-4 GPU per node, but can't provide you any feedback yet.


What is your main application for parallel rendering? What kind of
a setup you are interested in?


Best regards,

Makhinya Maxim

I am using equalizer for remote rendering of complex models.
I am interested in a setup with multiple nodes, one gpu each.

Thanks for your reply.

Regards,
Adam

[Stefan Eilemann]

Hi Adam,


On Wed, May 6, 2009 at 2:18 PM, gr3go (via Nabble)
<[hidden email]> wrote:
> Did anyone tested equalizer on bigger clusters like 10+ nodes?
> I don't have the possibility to benchmark on more nodes, so I would
> appreciate if anyone could provide some info related to this issue.

We've tested Equalizer on 10+ nodes. The speedup depends on your base
performance, data size and structure as well as network speed. As a
rule of thumb, 10-20 FPS are the upper limit for rendering 2k
(1920x1200) resolutions. Divide or multiply this number with your view
size.

We are working on pushing the boundary further, by optimizing the
pipeline and better compression. You're welcome to contribute here!


Cheers,

Stefan.

[gr3go]

Hi Stefan,

Stefan Eilemann wrote:

We are working on pushing the boundary further, by optimizing the
pipeline and better compression. You're welcome to contribute here!

What is the status of this issue? I checked the svn version and it seems it still uses RLE.

Cheers,

Adam.

[Stefan Eilemann]

On 09.05.2009, at 09:54, gr3go <[hidden email]> wrote:

>
> Hi Stefan,
>
>
> Stefan Eilemann wrote:
>>
>> We are working on pushing the boundary further, by optimizing the
>> pipeline and better compression. You're welcome to contribute here!
>>
>
> What is the status of this issue? I checked the svn version and it  
> seems it
> still uses RLE.

We are in the process of implementing a plugin interface, and will  
release at least one additional compressor with it. The API is  
documented on the website.


Best,

Stefan.

>
>
> Cheers,
>
> Adam.
>
>
>
> --
> View this message in context: http://n2.nabble.com/performance-measurement-tp2813299p2849273.html
> Sent from the Equalizer - Parallel Rendering mailing list archive at  
> Nabble.com.
>
>
> _______________________________________________
> eq-dev mailing list
> [hidden email]
> http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
> http://www.equalizergraphics.com

_______________________________________________
eq-dev mailing list
[hidden email]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

[gr3go]

Stefan Eilemann wrote:

We are in the process of implementing a plugin interface, and will  
release at least one additional compressor with it. The API is  
documented on the website.


Best,

Stefan.

Hy Stefan,

I took a look on the API, seem ok :)
When are you planning to release this compression plugin API?

Regards,
Adam

[Stefan Eilemann]

Hi Adam,

On 9. May 2009, at 21:38, gr3go (via Nabble) wrote:

> I took a look on the API, seem ok :)
> When are you planning to release this compression plugin API?

When it's ready. Finally I got to say this sentence. ;)

We are implementing it right now, it will be usable in a couple of  
weeks max.


Cheers,

Stefan.

[gr3go]

Stefan Eilemann wrote:

Hi Adam,

When it's ready. Finally I got to say this sentence. ;)

We are implementing it right now, it will be usable in a couple of  
weeks max.


Cheers,

Stefan.

Hi Stefan,

Please let us know when the compression API will be usable.

Best wishes,
Adam

[gr3go]

Hey,

I'v just tested equalizer performance on 11 PC-s (1 assambler node, 10 rendering)
For 2D and DB configurations the results seems to be acceptable.
For DPlex, the results are very strange, the performance is not increasing by increasing the number of stations...
Maybe the config files are wrong somehow...
Here
 is a config file for 8 rendering nodes.

Any hits?
Btw, I am using an application based on EqPly....

Regards,
Adam

[Stefan Eilemann]

On 14. Jun 2009, at 16:57, gr3go wrote:

> For 2D and DB configurations the results seems to be acceptable.
> For DPlex, the results are very strange, the performance is not  
> increasing
> by increasing the number of stations...

On a first glance, the config looks ok. Can you post a screenshot of  
the statistics?


Best,

Stefan.

_______________________________________________
eq-dev mailing list
[hidden email]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

[gr3go]

Stefan Eilemann wrote:

On 14. Jun 2009, at 16:57, gr3go wrote:

> For 2D and DB configurations the results seems to be acceptable.
> For DPlex, the results are very strange, the performance is not  
> increasing
> by increasing the number of stations...

On a first glance, the config looks ok. Can you post a screenshot of  
the statistics?


Best,

Stefan.

_______________________________________________
eq-dev mailing list
[e-mail]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

Hey,

Thanks for your quick response, Stefan.
Here is a screenshot for a 10 node DPlex config.

The initialization took a long time...somewhere around 3 mins....  With other configs it's moving lot more faster.Any ideas why?
The model has ~180mb, 5.000.000 vertices and 10.000.000 faces.

Regards,
Adam

[Stefan Eilemann]

On 14. Jun 2009, at 17:29, gr3go wrote:

> Thanks for your quick response, Stefan.
> http://img23.imageshack.us/img23/6971/dplex10nodess.jpg Here  is a
> screenshot for a 10 node DPlex config.

For some reasons your DPlex rendering don't overlap. Which version are  
you using?

I've just tested with trunk and 5-window.DPlex and everything look ok,  
see attached screenshot.


> The initialization took a long time...somewhere around 3 mins....  
> With
> other configs it's moving lot more faster.Any ideas why?

My guess would be that since your draw don't overlap, the  
serialization and display list creation in the first draw of each pipe  
is also serialized, thus taking 10x longer than normally.


Cheers,

Stefan.




_______________________________________________
eq-dev mailing list
[hidden email]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

[gr3go]

Stefan Eilemann wrote:

On 14. Jun 2009, at 17:29, gr3go wrote:

> Thanks for your quick response, Stefan.
> http://img23.imageshack.us/img23/6971/dplex10nodess.jpg Here  is a
> screenshot for a 10 node DPlex config.

For some reasons your DPlex rendering don't overlap. Which version are  
you using?

I've just tested with trunk and 5-window.DPlex and everything look ok,  
see attached screenshot.


> The initialization took a long time...somewhere around 3 mins....  
> With
> other configs it's moving lot more faster.Any ideas why?

My guess would be that since your draw don't overlap, the  
serialization and display list creation in the first draw of each pipe  
is also serialized, thus taking 10x longer than normally.


Cheers,

Stefan.



 
_______________________________________________
eq-dev mailing list
[e-mail]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

Hey,

It's my fault, I tested with clean svn build and the results are ok.
Thanks for the tips.

Regards,
Adam

[gr3go]

gr3go wrote:

Stefan Eilemann wrote:

On 14. Jun 2009, at 17:29, gr3go wrote:

> Thanks for your quick response, Stefan.
> http://img23.imageshack.us/img23/6971/dplex10nodess.jpg Here  is a
> screenshot for a 10 node DPlex config.

For some reasons your DPlex rendering don't overlap. Which version are  
you using?

I've just tested with trunk and 5-window.DPlex and everything look ok,  
see attached screenshot.


> The initialization took a long time...somewhere around 3 mins....  
> With
> other configs it's moving lot more faster.Any ideas why?

My guess would be that since your draw don't overlap, the  
serialization and display list creation in the first draw of each pipe  
is also serialized, thus taking 10x longer than normally.


Cheers,

Stefan.



 
_______________________________________________
eq-dev mailing list
[e-mail]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

Hey,

It's my fault, I tested with clean svn build and the results are ok.
Thanks for the tips.

Regards,
Adam

Hey,

I observed some interesting things: if the scene is rotation with a continous speed, the performance is increased (from 110fps to 180fps).
In case of a static scene, in my version, the system was rendering too (I commented the block starting with "while( !config->needsRedraw( ))" in eqPly).

What could be the reason for this to happen?

Regards,
Adam

[Stefan Eilemann]

On 16.06.2009, at 11:17, gr3go <[hidden email]> wrote:

>
>>
>>
>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>
>>
>>
>>
>>
>
> Hey,
>
> I observed some interesting things: if the scene is rotation with a
> continous speed, the performance is increased (from 110fps to 180fps).
> In case of a static scene, in my version, the system was rendering  
> too (I
> commented the block starting with "while( !config->needsRedraw( ))" in
> eqPly).
>
> What could be the reason for this to happen?

At this frame rate, it could be the framerate equalizer. It has to  
slow down the destination window occasionally to create a smooth  
swapbuffer rate.

Try disabling it to see if it is the reason.


Best,

Stefan.

>
>
> Regards,
> Adam
>
>
> --
> View this message in context: http://n2.nabble.com/performance-measurement-tp2813299p3085289.html
> Sent from the Equalizer - Parallel Rendering mailing list archive at  
> Nabble.com.
>
>
> _______________________________________________
> eq-dev mailing list
> [hidden email]
> http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
> http://www.equalizergraphics.com

_______________________________________________
eq-dev mailing list
[hidden email]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

[gr3go]

Stefan Eilemann wrote:

On 16.06.2009, at 11:17, gr3go <[e-mail]> wrote:
>
>>
>>
>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>
>>
>>
>>
>>
>
> Hey,
>
> I observed some interesting things: if the scene is rotation with a
> continous speed, the performance is increased (from 110fps to 180fps).
> In case of a static scene, in my version, the system was rendering  
> too (I
> commented the block starting with "while( !config->needsRedraw( ))" in
> eqPly).
>
> What could be the reason for this to happen?

At this frame rate, it could be the framerate equalizer. It has to  
slow down the destination window occasionally to create a smooth  
swapbuffer rate.

Try disabling it to see if it is the reason.


Best,

Stefan.
>
>
> Regards,
> Adam
>
>
> --
> View this message in context: http://n2.nabble.com/performance-measurement-tp2813299p3085289.html
> Sent from the Equalizer - Parallel Rendering mailing list archive at  
> Nabble.com.
>
>
> _______________________________________________
> eq-dev mailing list
> [e-mail]
> http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
> http://www.equalizergraphics.com

_______________________________________________
eq-dev mailing list
[e-mail]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com

Hey,

The same happens with 2D configs.
In the config file the framerate equalizer is not configured, and I suppose that by default is not enabled...Or am I wrong?

Adam

[Stefan Eilemann]

On 19. Jun 2009, at 0:07, gr3go wrote:

[180 FPS when rotating, 110fps when still in eqPly]

> The same happens with 2D configs.
> In the config file the framerate equalizer is not configured, and I  
> suppose
> that by default is not enabled...Or am I wrong?

Correct - the reason must be something else then. Either you  
investigate yourself where the time difference happens, or open a bug  
so we'll track this down later.


Cheers,

Stefan.


_______________________________________________
eq-dev mailing list
[hidden email]
http://www.equalizergraphics.com/cgi-bin/mailman/listinfo/eq-dev
http://www.equalizergraphics.com