[gmx-users] Regarding Gromacs 5.0.3 parallel computation

[Téletchéa Stéphane]

Le 10/12/2014 12:28, Bikash Ranjan Sahoo a écrit :
> Dear Dr. Stéphane,
>  Thank you for your quick reply. How can I solve this? Can you please 
> guide me in getting the right command to run mdrun as fast as gromacs 
> 4.5.5 -nt command. I am pasting the architecture of my cluster below. 
> Kindly help me understanding how can I modify the mdrun_mpi command to 
> access multi cores in my cluster.
>

Dear Bikash,

Why not using the -nt option in gromacs 5.0.3 too?
My point was that you should use the same parameters for comparison of 
performance...

What I would do is:

mdrun -nt 30 -pin auto

Try first using the gromacs tools (-pin auto) than using the "dplace" 
command since in your case
you ask it to split you job on different cpu according to the diagram 
you showed. At least you could
use the logical architecture for the dplace command:
http://techpubs.sgi.com/library/tpl/cgi-bin/getdoc.cgi?coll=linux&db=bks&srch=&fname=/SGI_Developer/LX_AppTune/sgi_html/ch05.html

Last, you should also try to first tune the pme decomposition using the 
"tunepme" option on a short run
to see if the rather rough domain decomposition done by mdrun in a first 
approach is optimal for your system.

You should also try to use cpu multiplicity and power of two for maximal 
performance,
in your case, probably something like -nt 12, -nt 24 or -nt 36, since 
each cpu seems to be a 12-core ...

See their respective manuals and command line helps for more info.

Best,

Stéphane

-- 
Team Protein Design In Silico
UFIP, UMR 6286 CNRS,
UFR Sciences et Techniques,
2, rue de la Houssinière, Bât. 25,
44322 Nantes cedex 03, France
Tél : +33 251 125 636
Fax : +33 251 125 632
http://www.ufip.univ-nantes.fr/ - http://www.steletch.org