[gmx-developers] (no subject)

[Erik Lindahl]

On Tuesday, Mar 4, 2003, at 23:02 US/Pacific, Peter Friedel wrote:

> dear gmx-developers,
> the last months I dealed successfully with the job to make GROMACS 
> available in multiprozessor mode. Now, we found out, that there are 
> different results using a different number of prozessors when running 
> the same job (the demo, which is given with the GROMACS package). the 
> package was compiled in --enable-float mode with lam-mpi on a 
> double-pentium IV processor machine (1600 MHz). we do not know, what 
> is the reason for and if this does or does not matter anything.
> could you please help us?

Hi Peter,

Some difference is quite normal; since you are summing forces and 
potentials in different order when running in parallel there will be 
slight numerical differences in the results. MD is a chaotic process, 
so this will lead to diverging trajectories, but the averages at 
equilibrium should be the same. Unless you are seeing very strange 
things in your resulting analysis I wouldn't worry.

To test it you should run a single step (or even zero steps to just 
calculate energies) and compare energies - those differences should be 
small, and even smaller in double precision.

Cheers,

Erik