Sounds to me like processes are not getting started correctly, which could
be an mpi installation issue. You have to understand your hardware, install
the appropriate mpi software, and link to the mpi libraries. You then have
to be sure of various issues regarding the hardware setup. There are issues
about dual cpu's, network cards, network switches, setting environment
variables and operating system parameters. All this stuff matters, and it
will affect pmemd differently than sander because pmemd uses more asynch mpi
functionality. BUT pmemd should be at least half again as fast as sander on
a small ethernet setup, if all the other stuff is setup correctly. I have
written lots of stuff out on the amber.scripps.edu webpage about this sort
of stuff; please take a look. Ross has also written some stuff I believe.
Best Regards - Bob Duke
----- Original Message -----
From: "Alessandro Nascimento" <al.s.nascimento.gmail.com>
To: <amber.scripps.edu>
Sent: Thursday, June 28, 2007 12:07 PM
Subject: AMBER: Improving pmemd parallel scaling
> Hi amber users,
>
> I'm running amber9 (lam-mpi) on 6 Intel Xeon 3.4 processors in Gigabit
> network.
> Looking the processors usage with linux "top" over the machines when
> running pmemd in a equilibrium simulation (~58K atoms, periodic box) I
> see that most of the times one processor is fully used, another one
> with ~50% usage and the resting are idle.
> sande.MPI has given me better results, although previous e-mail in
> this list argue that pmemd should scale better.
>
> My question is: which simulation parameters can be adjusted in order
> to achieve a better scaling? How can I improve it? Could there be any
> (hardware?) reason for the poor scaling in pmemd?
>
>
> Thanks in advance,
>
> --
> [ ]s
>
> -alessandro
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Jul 01 2007 - 06:07:26 PDT
