This is not possible today and there are no immediate plans to do so,  
but something like this could probably be supported with some work:

$ salloc --ntasks-per-node=0 -N8 --gres=gpu:1 bash


Quoting E L <[email protected]>:

> Hello,
>
> We have a GPU enabled slurm cluster. We are running OpenCL applications
> using VCL, a cluster platform that allows unmodified OpenCL applications to
> transparently utilize many OpenCL devices in a cluster,
> as if all the devices are on the local computer. This way of running
> applications requires allocating resources in a bit different way.
> For example, if we require 10 GPUs we need to allocate them on different
> computers, preferably without catching the CPUs as well.
>
> We are trying to find solution that will support both this new allocation
> scheme while allowing other users to still run gpu enabled MPI processes.
> What will be the best way to do it?
> Are there plans to decouple the GPU/CPU allocation?
>
> Thanks,
>
> nakee
>

Reply via email to