[gmx-users] parallel computations

[Berk Hess]

>From: fabio affinito <fabio.affinito at tiscali.it>
>Reply-To: Discussion list for GROMACS users <gmx-users at gromacs.org>
>To: Discussion list for GROMACS users <gmx-users at gromacs.org>
>Subject: Re: [gmx-users] parallel computations
>Date: Tue, 13 Jul 2004 11:49:09 +0200
>
>I tried shorter runs, and I found out that parallel computations seems to 
>be slower than serial...
>I.e.:
>               (Mnbf/s)   (MFlops) (ps/NODE hour) (NODE hour/ns)
>serial:
>Performance:      8.449    670.583      5.217    191.678
>np=1 MPI:
>Performance:      8.406    667.170      5.191    192.658
>np=2 MPI:
>Performance:      1.882    154.785      1.162    860.278
>
>Looking inside one of the logfiles of the last simulation I found:
>
>Total NODE time on node 1: 1268.41
>Average NODE time: 271.69
>Load imbalance reduced performance to 200% of max
>
>What does it means?

I guess you just have one molecule in your system?
The current gromacs version can not split a molecule
with constraints over different processors.
So all work is done on node 0 and using more nodes
only makes it slower.

Berk.

_________________________________________________________________
Play online games with your friends with MSN Messenger 
http://messenger.msn.nl/