[gmx-users] problems with parallel mdrun

[singh]

Dear Mark,

As you suggested, i ran the simulations with fourier spacing of
0.08,0.09,0.10 & 0.11 with mpirun -np 4 mdrun_d_mpi -np 4

following is the output

fourierspacing =0.08
-------------------------------------------------------
Program mdrun_d_mpi, VERSION 3.3.99_development_20070720
Source code file: gmx_parallel_3dfft.c, line: 90

Fatal error:
nx (63) and ny (63) must be divisible by the number of nodes (4).
-------------------------------------------------------

corresponding values from tpr files are
      nkx                  = 63
      nky                  = 63
      nkz                  = 189


fourierspacing =0.09 ; no errors  in simulations
corresponding values from tpr files are

      nkx                  = 56
      nky                  = 56
      nkz                  = 168

fourierspacing =0.10 ;
-------------------------------------------------------
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).
-------------------------------------------------------
corresponding values from tpr files are
      nkx                  = 50
      nky                  = 50
      nkz                  = 150



fourier spacing 0.11 ; no errors in simulations

corresponding values from tpr files are
     nkx                  = 48
      nky                  = 48
      nkz                  = 140


As you mentioned, the error does not occur when nx is divisible by number of
nodes

I have updated and compiled the CVS version yesterday only.


With Regards,
Gurpreet Singh

-------------------------------------------------
University of Dortmund
Department of Chemistry
Physical Chemistry I  -  Biophysical Chemistry
Otto-Hahn Str. 6
D-44227 Dortmund
Germany
Office:   C1-06 room 176
Phone:  +49 231 755 3916
Fax:     +49 231 755 3901
-------------------------------------------------

-----Original Message-----
From: Mark Abraham [mailto:Mark.Abraham at anu.edu.au] 
Sent: Thursday, August 09, 2007 10:01 AM
To: Discussion list for GROMACS users
Subject: Re: [gmx-users] problems with parallel mdrun

singh wrote:
> Dear Mark,
> Thanks for the reply. 
> I am using the following options for the electrostatics in the mdp file
> coulombtype         =  pme
> fourierspacing      =  0.1
> pmeorder            =  4
> optimize_fft        =  yes
> ewald_rtol          =  1.0e-5
> ewald_geometry      =  3dc
> 
> So if I understood you correctly, I have to change either fourier spacing
or
> my box size in order to get it running with -np 4

Hmm. If that's all the options relevant to electrostatics in your .mdp 
file, then this looks weird enough to be a possible bug, given that 
you're using a development version of mdrun.

First, update your CVS source, check this problem still happens. Then 
try it for a small range of fourierspacing values around 0.1 and see 
what behaviour you get. I suggest 0.08, 0.09, 0.1, 0.11, 0.12. If you 
report what you find, please use gmxdump on the input .tpr to report on 
nx/ny/nz.

On what I've seen so far, the automatic "choose nx/ny/nz suitable for 
grid spacing and number of nodes" might be broken. With this test, we're 
looking to see whether only choices of fourierspacing that produce 
nx/ny/nz that are accidentally divisible by the number of nodes are the 
only successful ones... The good news is the fix will be easy if this is 
the problem.

Mark