[gmx-developers] gromacs memory footprint
hess at kth.se
Fri Aug 8 10:08:35 CEST 2014
I think that old estimate of mine still shouldn't be too far off.
Why do you want to know this? Gromacs needs very little memory, in
practice you will never run into memory limits. We are more interested
in when things fit into L1/2/3 cache, but for that the full memory
footprint is not relevant.
On 08/08/2014 09:07 AM, Dimitris Dellis wrote:
> Is there any way to estimate the required memory per proc for MPI runs
> as function of atoms, cores etc ?
> (aware of the estimation here :
> For a system of ~120k atoms with PME, gromacs-4.6 pure MPI, verlet, rc
> 1.2 nm,
> the memory footprint of all except master procs is nearly constant
> when running with 16-1024 cores/1-64 nodes, (70 -->54 MB/proc).
> In a ~3M atoms system with RF, memory/proc decreases to ~half (43%)
> going from 16 to 2048 cores.
More information about the gromacs.org_gmx-developers