[gmx-users] Large Periodic Molecule

Rogan Carr rogan at ks.uiuc.edu
Tue Nov 30 21:16:28 CET 2010


Hi all,

I am having a problem, where my simulation fails when I run on  
multiple nodes.  On an 8-core machine, my job runs fine, but when I  
run on 240 processors (20 12-core nodes), I get the errors

"U-B of 59312 missing 1"

followed by

"Software inconsistency error:
Some interactions seem to be assigned multiple times"

I can use the .tpr file from the supercomputer (the 240 core job) and  
it runs fine on my 8-core machine as well.

I believe the problem is that I am simulating a system containing an  
"infinite" slab of amorphous silica, which is one large molecule,  
bonded to itself across the periodic boundaries.  I am using the  
charmm-compatible parameters developed by Cruz-Chu, Aksimentiev and  
Schulten (http://www.ks.uiuc.edu/Research/silica/SiO2ff.html), so I  
have each atom as its own charge group, and each atom is harmonically  
restrained to its initial coordinates (I also added in angle terms  
with small force constants so that I could use the GBSA solver, which  
requires all bonded triplets to have an angle defined).

 From my discussions with others, I have heard that it is common to  
have such problems with "infinite" molecules, where not all  
information about bonded interactions are shared between nodes.  One  
solution might be to use the old "particle decomposition"  
parallelization, but I would like to use domain decomposition for its  
speed.

Has anyone had any experience with this?

Thanks,
Rogan






More information about the gromacs.org_gmx-users mailing list