[gmx-developers] parallelization and accuracy
Anton Feenstra
feenstra at chem.vu.nl
Mon Feb 10 11:06:04 CET 2003
Frauke Meyer wrote:
>
> Dear all, dear Justin,
>
> I tried to change the pull code of the 3.1.4 version to enable it to
> run in parallel. In principle it works, but the force I obtain starts to
> differ (by 10^-3 kJ/mol/nm) from a sequential run from the 5th integration
> step on. The force profiles of the seq. and parallel run finally diverge
> after several 100 steps.
> Any hint, what the reason might be for that? The only things I changed are
> to add the stuff for parallel runs to the list of variables the procedures
> of the pull code are called with (start, homenr, cr) and to restrict I/O
> to the master node (actually I just transferred all changes for parallel
> pulling of the cvs version into my code).
Guessing from my intuition, without much factual basis, I might
suggest this is simply the result of rounding. It seems likely
that in paralell, some calculations are performed in different
order w/r to a sequential run, which would lead to different rounding
and hence to divergence in the results. IMHO, nothing to worry about,
as long as the first steps are, with execption of the last decimal,
identical.
--
Groetjes,
Anton
________ ___________________________________________________________
| | Anton Feenstra |
| . | Dept. of Pharmacochemistry - Vrije Universiteit Amsterdam |
| |---- | De Boelelaan 1083 - 1081 HV Amsterdam - The Netherlands |
| |---- | Tel: +31 20 44 47608 - Fax: +31 20 44 47610 |
| ' __ | Feenstra at chem.vu.nl - http://www.few.vu.nl/~feenstra/ |
| / \ |-----------------------------------------------------------|
| ( ) | Dept. of Biophysical Chemistry - University of Groningen |
| \__/ | Nijenborgh 4 - 9747 AG Groningen - The Netherlands |
| __ | Tel +31 50 363 4327 - Fax +31 50 363 4800 |
| / \ | K.A.Feenstra at chem.rug.nl - http://md.chem.rug.nl/~anton |
| ( ) |-----------------------------------------------------------|
| \__/ | "Jesus Built My Hotrod" (Ministry) |
|________|___________________________________________________________|
More information about the gromacs.org_gmx-developers
mailing list