GPU computing in spmd

Hi, I have two computers with cuda enable gpus when I run "gpuArray" locally, it workes well but when I run it in SPMD statement on 2 computers it shows this error: " No device supporting CUDA was found." if I want to have a cluster of GPUs what can I do? Thanks

[Edric Ellis]

What OS are cluster machines running? What GPUs and drivers are installed? What scheduler are you using?

[Aseman]

OS:64 bit Win7

GPU:GeForce GT 430

Driver:285.62-desktop-win7-64bit

scheduler:Jobmanager

Is it possible to use a GPU command in spmd on cluster?

Answer 1

[Jason Ross]

This should work, since it looks like that card supports TCC mode.

You can check (and change) this with the nvidia-smi utility, which should be in with the files that came with the driver in a directory called "NVSMI"

[Aseman]

Thanks for your answers

whould you please help me more ,what should I exactly do?

[Jason Ross]

c:\Program Files\NVIDIA Corporation\NVSMI>nvidia-smi --query

In the output, you'll see something like the following for your GT430:

GPU 0:8:0

Product Name : Tesla C1060

Display Mode : Disabled

Persistence Mode : N/A

Driver Model

Current : TCC

Pending : TCC

Then you can set the mode to TCC using the following. You can get the gpu id from the --query output

nvidia-smi --id=<gpu id> --driver-model=TCC

The GPU will now be in TCC mode and used for computation, and should be available.

[Jason Ross]

A small clarification: The output above is abbreviated, and the actual information will be different for a GT430, but you should end up in the "TCC" state. I'm guessing it will be in the "WDDM" driver model.

[Aseman]

Thanks for your answer I will check it and feedback but I think TCC mode is a feature of tesla cards.

[Jason Ross]

Yes, that looks to be the case with this card on Windows:

http://nvidia.custhelp.com/app/answers/detail/a_id/2595/kw/tcc/session/L3RpbWUvMTMyNDMzMDA1NC9zaWQvR0pDaUpfTGs%3D

Answer 2

[Friedrich]

Hi,

the doc says the following:

Access from a MATLAB worker running on a Microsoft Windows operating system with a job manager as the scheduler, requires an NVIDIA Tesla Compute Cluster (TCC) driver with an NVIDIA Tesla card.

http://www.mathworks.com/help/releases/R2011b/toolbox/distcomp/bsic3by-1.html#bsiqdwh