[gmx-users] parallel run crashes only when shuffled

[Erik Lindahl]

Mikko Huhtala wrote:

>David van der Spoel writes:
> > I forgot to add, yes shuffle coordinates manually and topology manually.
>
>How does this add to what you wrote before? :-)
>
>I guess the question is what does 'shuffle manually' mean? How does
>grompp decide what to put on which processor? It reads the topology
>file in the order it is written and divides the system in chunks of
>approximately equal size (without breaking molecules), correct? So
>manual shuffling means reordering the lines in the [ molecules ]
>section of the topology and making the coordinate file match.
>
Yes - I've done that quite a lot for lipid bilayers when running in 
parallel.

Assume you have 256 DPPC lipids and 100 waters per lipid (just an 
example, 100 is way too much).
The first time I run a new system I use the -shuffle option. The 
resulting .gro file will then be shuffled.

Instead of keeping on deshuffling and reshuffling the files, I just 
created a new topology
where the molecules section would look like (for 4 processors)


DPPC            64
SOL         6400
DPPC            64
SOL         6400
DPPC            64
SOL         6400
DPPC            64
SOL         6400

This order is the same as the shuffled .gro file, so then you can just 
forget about the shuffling. grompp
won't split a molecule over processors if you are constraining ALL 
bonds, and if you're not constraining
all bonds it's quite OK to split it...

Cheers,

Erik