Chris -
Thanks for tracking this down. I think the crucial difference between
sander and pmemd that caused the susceptibility is likely that pmemd makes a
lot more use of asynchronous i/o in mpi, thus enabling overlap of
computation and data communications. I did not pick up the fact that there
was a myrinet patch because by the time we were running ifc 7/7.1, we were
already running a later release of the myrinet software (or perhaps my
systems folks here just did it silently in the background).
Regards - Bob
----- Original Message -----
From: "Chris Moth" <Chris.Moth.vanderbilt.edu>
To: <amber.scripps.edu>
Cc: "Robert Duke" <rduke.email.unc.edu>
Sent: Monday, March 29, 2004 2:05 PM
Subject: AMBER: PMEMD and myrinet trouble
> At 03:56 PM 3/24/2004, Robert Duke wrote:
> >Chris -
> >I would first suspect the hardware, based on my past experience with
> >myrinet.
>
> Robert:
>
> Thanks for all your round-the-clock help and suggestions with our
> PMEMD-on-Myrinet troubles. All is now solved.
>
> I hope our notes below might save time for other groups in future.
>
> As you know, we initially suspected that our bad minimization and MD
> trajectory data was a software problem with PMEMD - primarily because
> SANDER ran fine on our myrinet
> cluster: http://structbio_dot_vanderbilt.edu/comp/hw/piranha/
>
> That said, PMEMD always has run well on all our other platforms.
>
> It turns out that there was a myrinet software problem on our end.
>
> Installing the latest mpich software from
> http://www.myri.com/scs/#downloads was the solution. This update is
> required for multi-threaded software (which presumably SANDER is not)
built
> with ifc 7.1.
>
> We've also learned the importance of regression testing with PMEMD in
> addition to SANDER as we install upgrades to our platform (hardware, OS,
or
> drivers).
>
> Small incompatibilities with platform upgrades can cause major trajectory
> failures, without warning or error messages!
>
> Again, thanks for all our help double-checking things on your hardware.
>
> Sincerely
>
> Chris Moth
>
>
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Mon Mar 29 2004 - 20:53:01 PST
