[gmx-users] possible configuration for gromacs gpu compute node

tsikosek t.sikosek at utoronto.ca
Tue Apr 8 18:23:08 CEST 2014


Hi,

you can use the mdrun options to reserve specific CPU cores and GPU cards
for that run. E.g. if you have two GPUs they will be indexed 0 and 1 so you
just say "-gpu_id 0" for one simulation and "1" for the other. For the CPUs
it should be sufficient to specify the number of OpenMP threads per run via
"-nt 6".

Whether or not 6 CPUs is sufficient for your simulation depends on the
system size. My guess is that is that it should be ok, especially with the
GPU. But you can just do a benchmark where you run the actual production
simulation for just a few thousand time steps using mdrun's "-nsteps"
option. You can then choose different numbers of OpenMP threads and see how
your performance (ns/day) changes. There will be some optimal number of CPUs
for your specific system that you can find out. You can also play with
domain decomposition.

Hope that helps.

Tobias

--
View this message in context: http://gromacs.5086.x6.nabble.com/possible-configuration-for-gromacs-gpu-compute-node-tp5015720p5015743.html
Sent from the GROMACS Users Forum mailing list archive at Nabble.com.


More information about the gromacs.org_gmx-users mailing list