How to convert single-GPU code to double-GPU ?

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi,

I am very new to CUDA programming. I have an opensource code, DualSPHysics, which can run only on single-GPU. DualSPHysics is written in C++ and the CPU version can run only on single-node (OpenMP) and the GPU version only on single-GPU (CUDA). But, the node I use for computing has two GPU cards (K20’s).

I want to convert this DualSPHysics code to run on the two GPU’s. Please enlighten me on what should I do to make this happen.

Thank you very much,

Best regards,
Sunil

#2

Hello,

You can run 1 program with 2 cards, by using cudasetdevice and manually divide the work between the gpus using streams.

Cristian

#3

Hi Cristian,

Thank you very much for your reply.

Can you please point me to some online material which explains with examples how I can “manually divide the work between the gpus using streams” ?

Best Regards,
Sunil

#4

Hello,

I would suggest to read the book CUDA by Example https://developer.nvidia.com/content/cuda-example-introduction-general-purpose-gpu-programming-0

There are also good articles about streams on the nvidia Parallel for all blog. http://devblogs.nvidia.com/parallelforall/?s=streams&submit=Search

The basic ideea is to divide the work in 2 equal parts (if the cards are identical) have them perform the calcuations, then collect the data. The biggest problem is the communication between the cards.

#5

Maybe this will help:

http://users.wfu.edu/choss/CUDA/docs/Lecture%2013.pdf