Numerical Simulation and Visualization

[Andreas Kirsch (in.tum)]

Hello,
I want to write a PDE simulation and parallelize it using Equalizer. I
only have no idea how to go on about the exact design.

The application consists of two parts:
* simulation
* visualization
Simulation:
I have a volume of cells I need to update every frame (e.g. some fluid
or gas) using the GPU because of its parallel computing power. It's a
big cubic array of data and I'd like to split it up into sub cubes and
keep and update the data on different nodes.
That is every node contains a certain sub volume and the different nodes
only need to exchange the boundary data.

Visualization:
The visualization part would require volume rendering using a standard
approach ala eqEvolve to parallelize. Every node would contribute a
rendering of the data it has available locally.

I have no idea whether Equalizer offers any features for doing such
general purpose computing (that is, I haven't found any) and I'd like to
know whether there is already anything built in or if there is anything
planned.
Also I have difficulties imagining how to express both parts of the
application in the Node/Pipe/Window/Channel paradigm.

I also have a related question regarding the network code:
If I were to code the simulation part on my own, I'd need to send some
specific data to each node from each adjacent node. Using the current
network model my idea would be that each node has a "state" master
object and nodes that are interested in the node's state (it would
contain the boundary surface data for example) would subscribe to that
object and sync to it every time they need to access the data.
The question is how do I tell other nodes about my state node?
Is the only way using custom commands?
It would be nice if there was a way to do this transparently. I've seen
there exists a method called becomeMaster(), but it only updates the
former master object and not its slaves. My idea would have been to
create the state objects at the app node first, all nodes map the
general state object and each node then takes over ownership of its
corresponding state object, but this obviously will run into issues
using the current becomeMaster() method.

That's a lot of text, I hope I was able to communicate my problem(s). I
also hope that someone can help me with it and nudge me into the right
directions.
Also, if there is anything way too unclear, I'll be happy to elaborate
on it.

Thanks and cheers,
 Andreas Kirsch

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

[Philippe Robert]

Hi,

it's good that you bring this up. We indeed have concrete plans to  
extend Equalizer in such a way that GPU based simulations and other  
GPGPU applications can be integrated seamlessly. Until this is the  
case you have to do a bit more work by yourself, but it is still  
possible.

In your case you could integrate the compute tasks into the window's  
or channel's draw function. To distribute the updated data you then  
might want to use custom events and let the master send the relevant  
changes to the clients on a per frame basis. Does this make sense to  
you?

I am aware that this approach is probably not optimal, but it works.  
We will soon provide a simple example that outlines the basic idea.

BTW if there are others on this list who are interested in doing GPU  
computing using Eq, please let me know!

Cheers,
Philippe

Am Jun 30, 2009 um 22:11 schrieb "Andreas Kirsch (in.tum)" <[hidden email]
 >:

> Hello,
> I want to write a PDE simulation and parallelize it using Equalizer. I
> only have no idea how to go on about the exact design.
>
> The application consists of two parts:
> * simulation
> * visualization
> Simulation:
> I have a volume of cells I need to update every frame (e.g. some fluid
> or gas) using the GPU because of its parallel computing power. It's a
> big cubic array of data and I'd like to split it up into sub cubes and
> keep and update the data on different nodes.
> That is every node contains a certain sub volume and the different  
> nodes
> only need to exchange the boundary data.
>
> Visualization:
> The visualization part would require volume rendering using a standard
> approach ala eqEvolve to parallelize. Every node would contribute a
> rendering of the data it has available locally.
>
> I have no idea whether Equalizer offers any features for doing such
> general purpose computing (that is, I haven't found any) and I'd  
> like to
> know whether there is already anything built in or if there is  
> anything
> planned.
> Also I have difficulties imagining how to express both parts of the
> application in the Node/Pipe/Window/Channel paradigm.
>
> I also have a related question regarding the network code:
> If I were to code the simulation part on my own, I'd need to send some
> specific data to each node from each adjacent node. Using the current
> network model my idea would be that each node has a "state" master
> object and nodes that are interested in the node's state (it would
> contain the boundary surface data for example) would subscribe to that
> object and sync to it every time they need to access the data.
> The question is how do I tell other nodes about my state node?
> Is the only way using custom commands?
> It would be nice if there was a way to do this transparently. I've  
> seen
> there exists a method called becomeMaster(), but it only updates the
> former master object and not its slaves. My idea would have been to
> create the state objects at the app node first, all nodes map the
> general state object and each node then takes over ownership of its
> corresponding state object, but this obviously will run into issues
> using the current becomeMaster() method.
>
> That's a lot of text, I hope I was able to communicate my problem
> (s). I
> also hope that someone can help me with it and nudge me into the right
> directions.
> Also, if there is anything way too unclear, I'll be happy to elaborate
> on it.
>
> Thanks and cheers,
> Andreas Kirsch
>
> _______________________________________________
> 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

[Andreas Kirsch (in.tum)]

Hello Philippe,

Philippe Robert wrote:

> Hi,
>
> it's good that you bring this up. We indeed have concrete plans to  
> extend Equalizer in such a way that GPU based simulations and other  
> GPGPU applications can be integrated seamlessly. Until this is the  
> case you have to do a bit more work by yourself, but it is still  
> possible.
>
> In your case you could integrate the compute tasks into the window's  
> or channel's draw function. To distribute the updated data you then  
> might want to use custom events and let the master send the relevant  
> changes to the clients on a per frame basis. Does this make sense to  
> you?
>  

Yeah,  this sounds good. It would probably be better though to exchange
data directly between clients because otherwise the master might become
a bottleneck bandwidth-wise, but it should be sufficient for testing.
> I am aware that this approach is probably not optimal, but it works.  
> We will soon provide a simple example that outlines the basic idea.
>  
Neat, this sounds like a very good thing to do.
> BTW if there are others on this list who are interested in doing GPU  
> computing using Eq, please let me know!
>
> Cheers,
> Philippe
>  
In my opinion GPGPU support would be a major feature and supporting
things like load balancing inside the library would be a great selling
point, because that is generally a big PITA when working on
parallelizing projects and having a middle ware for that would be nifty.
Since support for FBOs and offscreen-rendering was added to Equalizer I
think one could naturally extend Equalizer's design from canvases and
frames, etc. to computing tasks and data dependencies that are more
complex than the currently supported ones.
Thanks for your quick and helpful reply.

Cheers,
Andreas

PS: Im resending this because there was a problem with my uni email
account and I couldnt find the message in the eqdev archive. My
apologies if you have already received it.

> Am Jun 30, 2009 um 22:11 schrieb "Andreas Kirsch (in.tum)" <[hidden email]
>  >:
>
>  
>> Hello,
>> I want to write a PDE simulation and parallelize it using Equalizer. I
>> only have no idea how to go on about the exact design.
>>
>> The application consists of two parts:
>> * simulation
>> * visualization
>> Simulation:
>> I have a volume of cells I need to update every frame (e.g. some fluid
>> or gas) using the GPU because of its parallel computing power. It's a
>> big cubic array of data and I'd like to split it up into sub cubes and
>> keep and update the data on different nodes.
>> That is every node contains a certain sub volume and the different  
>> nodes
>> only need to exchange the boundary data.
>>
>> Visualization:
>> The visualization part would require volume rendering using a standard
>> approach ala eqEvolve to parallelize. Every node would contribute a
>> rendering of the data it has available locally.
>>
>> I have no idea whether Equalizer offers any features for doing such
>> general purpose computing (that is, I haven't found any) and I'd  
>> like to
>> know whether there is already anything built in or if there is  
>> anything
>> planned.
>> Also I have difficulties imagining how to express both parts of the
>> application in the Node/Pipe/Window/Channel paradigm.
>>
>> I also have a related question regarding the network code:
>> If I were to code the simulation part on my own, I'd need to send some
>> specific data to each node from each adjacent node. Using the current
>> network model my idea would be that each node has a "state" master
>> object and nodes that are interested in the node's state (it would
>> contain the boundary surface data for example) would subscribe to that
>> object and sync to it every time they need to access the data.
>> The question is how do I tell other nodes about my state node?
>> Is the only way using custom commands?
>> It would be nice if there was a way to do this transparently. I've  
>> seen
>> there exists a method called becomeMaster(), but it only updates the
>> former master object and not its slaves. My idea would have been to
>> create the state objects at the app node first, all nodes map the
>> general state object and each node then takes over ownership of its
>> corresponding state object, but this obviously will run into issues
>> using the current becomeMaster() method.
>>
>> That's a lot of text, I hope I was able to communicate my problem
>> (s). I
>> also hope that someone can help me with it and nudge me into the right
>> directions.
>> Also, if there is anything way too unclear, I'll be happy to elaborate
>> on it.
>>
>> Thanks and cheers,
>> Andreas Kirsch
>>
>> _______________________________________________
>> 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
>
>  

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

[Philippe Robert]

Hi,

On Jul 3, 2009, at 6:40 PM, Andreas Kirsch (in.tum) wrote:

>> In your case you could integrate the compute tasks into the window's
>> or channel's draw function. To distribute the updated data you then
>> might want to use custom events and let the master send the relevant
>> changes to the clients on a per frame basis. Does this make sense to
>> you?
>>
> Yeah,  this sounds good. It would probably be better though to  
> exchange
> data directly between clients because otherwise the master might  
> become
> a bottleneck bandwidth-wise, but it should be sufficient for testing.

I have just committed a basic example which shows exactly how to  
achieve that using sort-first distribution; i.e. using point-to-point  
communication. It is a simple port of the nbody example from the CUDA  
SDK. Hence, it is by no means efficient because there are far too many  
memory transfers and GPU up/downloads involved, but it illustrates the  
strategy quite well I hope.

More to come...

cheers,
Philippe



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

[Andreas Kirsch (in.tum)]

Hello Philippe,
thank you very much. I'm going to take a look at it now.
Cheers,
 Andreas

Philippe C.D. Robert wrote:

> Hi,
>
> On Jul 3, 2009, at 6:40 PM, Andreas Kirsch (in.tum) wrote:
>  
>>> In your case you could integrate the compute tasks into the window's
>>> or channel's draw function. To distribute the updated data you then
>>> might want to use custom events and let the master send the relevant
>>> changes to the clients on a per frame basis. Does this make sense to
>>> you?
>>>
>>>      
>> Yeah,  this sounds good. It would probably be better though to  
>> exchange
>> data directly between clients because otherwise the master might  
>> become
>> a bottleneck bandwidth-wise, but it should be sufficient for testing.
>>    
>
> I have just committed a basic example which shows exactly how to  
> achieve that using sort-first distribution; i.e. using point-to-point  
> communication. It is a simple port of the nbody example from the CUDA  
> SDK. Hence, it is by no means efficient because there are far too many  
> memory transfers and GPU up/downloads involved, but it illustrates the  
> strategy quite well I hope.
>
> More to come...
>
> cheers,
> Philippe
>
>
>
> _______________________________________________
> 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