Hello,

I create an allocation with 3 tasks and 1 GPU using the following command:

salloc -n3 --gres=gpu:1

Note: I would like each task to use the same GPU.  I try running the following 
tasks, but they do not run simultaneously:

srun -n1 --gres=gpu:1 myProg -arg abc
srun -n1 --gres=gpu:1 myProg -arg def
srun -n1 --gres=gpu:1 myProg -arg ghi


Does anyone know if this is a viable approach?  Do you have a suggested or 
alternative solution?

I've read that multiple GPUs can be defined in the gres.conf, such that GPUs 
can be over-subscribed.  However, this is very application-specific, so I would 
like to avoid this solution.  That is, I know for my application, I can run 3 
jobs per GPU, but that might not always be the case - therefore, I would prefer 
the user to determine the usage depending on their application.


Thanks,
adam


Reply via email to