[gmx-users] Paralellization limit?

[Carsten Kutzner]

Hi Andrea,

Andrea Carotti wrote:
> Hi all,
> I'm trying to simulate a system with two identical proteins (42aa each 
> one), solvent spc (18430 mols) and 6 ions NA+...for a total of  ~56100 
> atoms.
> Now the "problem" is that if I run the MD on 4 nodes everything works 
> fine, but when I try to use 6 or 8 cpus the process stops on the master 
> and continue only on the last two slaves (4 cpus, the systems are dual 
> xeon). The problem is that I can't find any error message on the log files.
> I've tried using the d.dppc benchmark files, and all goes fine with 
> 2-4-6 and 8 cpus.
> So my question is: is there a limit of parallelization in gromacs 
> depending on the simulated system?
> I'm using gromacs 3.3.1, mpich 1.2.5.2 and fftw 2.1.5.

I have encountered similar problems when using mpich 1.2.x on
Ethernet. Upgrading to mpich-2.x or using LAM solved the problems.
There is only a parallelisation limit in the sense that using more
CPUs does not always result in faster execution. 56 k atoms should
run happily on 8 CPUs.

Good luck,
   Carsten