I would have said "try exporting CUDA_VISIBLE_DEVICES without the quotes,"
but that doesn't seem to be the problem when I try it. There's probably
something with the server itself--pmemd.cuda will happily start up on a
device that already has a job running (both jobs will just run at slightly
under 50% their normal speed, pretty much like a CPU, assuming that
collectively they're not exceeding the card's on-board memory capacity).
Also helpful to google your error:
https://devtalk.nvidia.com/default/topic/470566/cuda-programming-and-performance/why-quot-all-cuda-capable-devices-are-busy-or-unavailable-quot-/
Your fellow sufferers seem to have found a remedy in updating their OS,
which probably updated various drivers. It may also be that there are
settings made through nvidia-smi that are preventing you from accessing
that particular card, if such a numbered card exists at all (going by
Jerrano's reply).
Dave
On Fri, Jun 16, 2017 at 3:48 PM, Krantzman, Kristin D <KrantzmanK.cofc.edu>
wrote:
> Greetings Amber Users,
> I am trying to use pmemd.CUDA on different GPU processors using the export
> CUDA_VISIBLE_DEVICES command.
> However, when I type export CUDA_VISIBLE_DEVICES="1" and then try to
> execute pmemd.cuda, I get the error that all CUDA-capable devices are busy
> are unavailable. When I type nvidia-smi, it seems that only one of the
> three GPUS is in use. Does anyone have experience with solving this
> problem?
> Thank you in advance.
> Best, Kristin
> Kristin D. Krantzman
> Professor, Department of Chemistry and Biochemistry
> College of Charleston
> Charleston, SC 29424
> (843) 953-3378
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Jun 16 2017 - 13:30:04 PDT
