[gmx-users] problems with parallel mdrun

[Mark Abraham]

Carsten Kutzner wrote:
> Mark Abraham wrote:
>> Gurpreet Singh wrote:
>>> I get the following errors while using paralled version of mdrun compiled
>>> with openmpi.
>>> mpirun -np 4 mdrun_d_mpi -np 4 -v -deffnm EQUI1
>>> *****
>>> Program mdrun_d_mpi, VERSION 3.3.99_development_20070720
>>> Source code file: gmx_parallel_3dfft.c, line: 90
>>>
>>> Fatal error:
>>> nx (50) and ny (50) must be divisible by the number of nodes (4).
>>> *******
>>> I am using a cubic box of 5nm and grid spacing is 0.1. The simulations
>>> runs fine if i use 5 nodes
>>> I am using a machines with 2 dual xeon processors so -np 4 is more
>>> appropriate for my sytems.
>>>  Is there any way to use -np 4  without changing the gridspacing or
>>> box size?
>> No. I'm guessing your .mdp file is trying to specify nx/ny/nz, rather
>> than the grid spacing. If you specify the latter, grompp will take care
>> of this issue so that you get the largest allowable spacing smaller than
>>  (or equal to) the requested grid spacing.
>>
> Hi,
> 
> in the current CVS version of grompp you cannot use the -np switch
> anymore. You will have to comment out the 'fourierspacing = 0.1' line
> and set
> 
> fourier_nx               = 52
> fourier_ny               = 52
> fourier_nz               = 52
> 
> in order to get a tpr file that you can run on 4 processors.
> Alternatively, stick to the original mpd settings but use the 3.3.1
> version of grompp with -np 4. These tpr files are read by CVS gromacs.

Aha, that explains it. Gurpreet, don't bother with the testing procedure 
I just outlined. Try one of Carsten's suggestions.

Mark