On Wed, Dec 2, 2015 at 8:53 AM, Arjun Sharma <arjunsharma83.gmail.com>
wrote:
> Hi,
>
> I’m struggling to get GPU-accelerated explicit solvent pH-replica exchange
> jobs to run. I have access to 2 NVIDIA Tesla K20x GPU’s.
> Pmemd CUDA-enabled mpi build runs just fine for non-REMD runs and provides
> great acceleration to my simulations. However, this doesn’t work for my
> 12-replica pH-REMD runs. Does each replica needs its own GPU ? or can GPU
> be shared by multiple replicas ? Is there a way I can divide 2 GPU nodes
> available to run 12 replicas ?
>
As Adrian mentioned, the Amber code is optimized for dedicated GPUs. If
your GPU is not in exclusive-process mode, you *can* assign multiple
replicas for each GPU, but the throughput will be reduced by at least half
(if you have 2 replicas per GPU, for instance) -- likely more than that.
That said, don't use the GPU for explicit solvent CpHMD or pH-REMD
simulations with Amber 14. The performance is abysmal (you will do better
with a small handful of processors per replica using pmemd.MPI). Amber 16
(in theaters everywhere this April) will fix this performance issue. But
until then, explicit solvent CpHMD should be considered CPU-only.
HTH,
Jason
--
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Dec 02 2015 - 06:30:08 PST
