[gmx-users] Runs OK on small CPS but fails on large number of CPUs...

[Mark Abraham]

On 29/09/2011 12:44 PM, Tandia, Adama wrote:
> Dear Co-users,
>
> I have a systems made of simple molecules, KNO3 to say. A classical MD
> runs to completion on a supercomputer when using 24 CPUs only and fails
> after few hundreds of ps, not always at the same step, when I increase
> the number of CPUs to 72, 84, 96 ...
> "The error message is as follows: " 1 particles communicated to PME node
> 1 are more than 2/3 times the cut-off out of the domain decomposition
> cell of their charge group in dimension x. This usually means that your
> system is not well equilibrated."

A given simulation system and code cannot scale efficiently to an 
arbitrary number of processors. GROMACS needs about 500-1000 atoms per 
processor to get good value. Close to the lower limit, things become 
increasingly likely to break. However, it is likely that your system is 
not actually equilibrated.

>
> I do know for facts that the system is very well equilibrated since the
> system had previously been equilibrated with EM on another machine until
> the energy does not change any more.

Successful EM does not imply equilibration, or nobody would ever worry 
about equilibration once EM succeeded...

Mark

>
> Is there any brilliant idea out there?
>
> Thanks in advance for your help.
> Regards,
>
> Adama
>
> ______________________
>
> Dr. Adama Tandia | Modeling&  Simulation, Corning INC, Corning, NY 14831
> USA | Tel: (+1) 607 248 1036
>