[gmx-users] Domain decomposition problems

[Mark Abraham]

On 25/02/2012 4:52 AM, Olivia Waring wrote:
> Dear Gromacs users,
>
> Could someone please shed light on the following problem? My system 
> consists of an alkanethiol SAM (in the xy plane) with a layer of water 
> on top. The mdrun command works when the box vectors are 1.1, 1.3, and 
> 2.5, but the SAM flies apart and the simulation crashes. I'm trying to 
> lengthen the box in the z direction (and of course, resolvating the 
> system) to see if that solves the problem. Unfortunately, this seems 
> to give rise to domain decomposition issues... I get the following 
> output in my md.log file:
>
> Initial maximum inter charge-group distances:
>     two-body bonded interactions: 0.424 nm, LJ-14, atoms 36 39
>   multi-body bonded interactions: 0.424 nm, Ryckaert-Bell., atoms 36 39
> Minimum cell size due to bonded interactions: 0.467 nm
> Maximum distance for 5 constraints, at 120 deg. angles, all-trans: 
> 0.794 nm
> Estimated maximum distance required for P-LINCS: 0.794 nm
> This distance will limit the DD cell size, you can override this with 
> -rcon
> Scaling the initial minimum size with 1/0.8 (option -dds) = 1.25
> Optimizing the DD grid for 6 cells with a minimum initial size of 0.992 nm
> The maximum allowed number of cells is: X 1 Y 1 Z 5
> Fatal error:
> There is no domain decomposition for 6 nodes that is compatible with 
> the given box and a minimum cell size of 0.9925 nm
> Change the number of nodes or mdrun option -rcon or -dds or your LINCS 
> settings
>
> I've experimented a bit with -rcon and with the number of nodes, but I 
> really have no idea what I'm doing! Any thoughts would be appreciated.

A given implementation cannot run an arbitrary simulation system 
efficiently in parallel on an arbitrary number of processors. The 
GROMACS implementation needs to be able to localize a chunk of nearby 
atoms to the same processor in a 3D grid of processors. Your system is 
too small for it to find a satisfactory chunk for each processor. Use 
more atoms or fewer processors. A very rough guideline is to use at 
least about 1000 atoms per processor - but the real constraints are 
geometric.

Mark