Multiple GPU Issue

[cesnet]

Multiple GPU Issue

[MartinPulec]

Hi Tony,

I am not sure if I'll be able to reproduce the problem because I am not sure if am able to make a mac supporting CUDA with multiple GPUs. I've, however, tried to do so in Linux and it worked fine for me using UltraGrid:

xpulec@bunny:~$ nvidia-smi

| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0      3054      C   bin/uv                                       117MiB |
|    1      3109      C   bin/uv                                       117MiB |
+-----------------------------------------------------------------------------+

These are just 2 independent processes using the GPU with GPUJPEG.

Would you be able to provide a working example code presenting the problem? Or at least instructions how to replicate (eg. timing of the processes etc.).

Thanks,
Martin

EDIT: I've also tested Mac (although only 10.11) with NV 960+680 and it seemed to work.

[wolfviking0]

Hi Martin,

Here my configuration, currently I am using Python (I create a wrapper using swig). So I am creating the process using the multiprocessing API.
I will test with pycuda to see if It give me the same error when I try to initialize cuda from a subprocess.
I was able to use another approach, inside the main process python, I instantiate two GPUJPEG instance and then pass them to the process, now it seems to be working.

Unfortunately I still have one trouble, when I try to set the device_id to the GPU 1 it seems to be running on 0.
In order to choose the device I did 'gpujpeg_init_device(device_id, flags)' it should be enough ? I don't have to specify something else ?

[wolfviking0]

@MartinPulec There is something in my code for sure, If I test the gpujpeg binary I can use the other GPU when I use -D 1. I am still searching my mistake.

[wolfviking0]

So the issue was related to the multiprocessing in python. I was able to change in my code how python start a process and now I am able to run 4 instance of gpujpeg on two GPU. It work really well thanks