Device selection in a hetrogenous host network render

[ejanne]

I am considering buying indigo and have been experimenting with the network rendering features using the trial version. I'm using my MBP with a GT 650M as master, and a linux box with a Quadro 3800 FX and two Tesla 2050s as slave. I'm initiating the render from within C4D. The problem I'm having is that the slave defaults to the Quadro 3800FX. Is there a way to specify the render device on the slave?

[Juju]

I'm taking a wild stab here as I've never had opportunity to work with Tesla's. Have a look in the machine's BIOS options, you should be able to set the primary display device there, I'm not sure if this will work here but you can try that.

Also remember that Indigo doesn't do multi GPU at the moment.

[ejanne]

I'm taking a wild stab here as I've never had opportunity to work with Tesla's. Have a look in the machine's BIOS options, you should be able to set the primary display device there, I'm not sure if this will work here but you can try that.

Also remember that Indigo doesn't do multi GPU at the moment.

Thank you for the suggestion, I will try that, but I'd rather not use the Teslas as display devices, since I have the 3800 FX for that. I know it's possible to specify which device by submitting on the command-line, or by editing the igs file. That's relatively easy when dealing with a single frame. But when I'm submitting an animation it would be preferable to set the target device from the point of submission in C4D, or by setting a default in a config file somewhere on the slave.

The only other alternative I can think of is to write my own script to reprocess the igs files. That would be a bit of a pain, since the exporter doesn't seem to support relative resource and save paths, meaning that the system relies on the master to marshal the results and assemble and save the final image.

As for multi GPU: that is not as important to me. Multi-GPU I think would be primarily important for rendering highly complex single stills where GPU memory is a constraint. However, I would be using Indigo mainly to render animations with relatively simple geometry, so it would make more sense to distribute whole frames to each processor, rather than distributing each frame across all the processors. It's probably more efficient, too, because it avoids replication of data. Although the way the Indigo network rendering system doesn't seem to support this, I think it wouldn't be too hard to write a script that would distribute frames on the command-line, simultaneously to both Teslas, the 3800 FX, AND possibly even the CPU (provided I dedicate enough threads to keeping the GPU's fed). I wrote such a script to distribute Houdini renders across a network of workstations when I didn't have access to more sophisticated farm software ... admittedly a kludge, but it got the job done. I'm sure it would be much easier syncing between processes on a single box than across multiple workstations over the network.

[OnoSendai]

Hi ejanne,
The way it works currently is that the render slave will pick what it considers the best GPU device to use, based on the estimated performance of each local device.
However I think I will add some code so that if a particular GPU device is specified in the local (for each slave) inifile.xml, then it will use that device.
This change will be probably be available in the next Indigo (3.6.14).
I'll write some details on how to specify the GPU device in the inifile.xml when the change is released.

Cheers!

[ejanne]

Excellent! That would make things a lot easier.

In my crude tests, the Quadro 3800 FX faired rather poorly against the C2050, and even slower than rendering on CPU.

Is it possible to start multiple slave processes on a single box, each targeting a separate GPU? Or will Indigo try to split up each frame amongst those slaves rather than submitting successive frames to each slave?