Dear Dr. Case,
Thank you for your quick reply.
I tried to run my system with 15, 16, 24 and 48 cores, obtaining the same error message in each case. Using 14 cores or less is unfortunately not an option, given the low speed I observed so far with 14 cores for my system. I need to use the AMBER02 polarizable force field which is not implemented in pmemd, unfortunately; and the AMOEBA force field, which is implemented, would be too expensive for my system.
Thanks for the clarification regarding sander's residue-based distribution method, this was indeed unclear to me.
I will send you my sander files (topology, coordinates, mdin) in a separate email as I don't want to send big attachments to everyone in the mailing list. (I can of course send them to anyone who is also interested in taking a look.)
All the best,
Marian Breuer
________________________________________
From: David A Case [case.biomaps.rutgers.edu]
Sent: 01 November 2011 12:22
To: AMBER Mailing List
Subject: Re: [AMBER] Issue with running sander with SHAKE
On Mon, Oct 31, 2011, Breuer, Marian wrote:
>
> I'm currently trying to carry out MD simulations with sander (MPI
> parallel version, AMBER11) on a protein system using SHAKE bond
> constraints for bonds containing hydrogen (sander option ntc=2). These
> work with up to 14 processors, but with any number of cores higher than
> that the simulations abort in the first step with the error message:
>
> partition error in shake on processor 0
> this processor has atoms 1 through 8265
> atom 8265 is within this range
> atom 8266 is not within this range !
Does the code work with 16 processors (or 32, i.e. an even power of 2)? Code
with an even power of 2 has been tested a lot more than for non-powers of 2.
How many processor counts above 14 have you tried?
I know these are big files, but if you could somehow post the prmtop and
coordinate files involved here (add mdin as well) it would greatly accelerate
finding what is wrong.
As workarounds: can you use pmemd instead of sander? It will be faster
anyway. Or, if you are on sander, live with 14 processors or fewer? (Sander
may not scale all that well beyond a dozen threads anyway.
> The total number of atoms in my system is 122321, which,
> when divided by 8265, yields roughly 14.8. This seems to suggest that
> sander always places 8265 atoms on each of the first 14 cores as long as
> there are more than 14 cores....
This inference is not valid. Sander uses a residue-based distribution method,
which may have some corner cases that fail.
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Nov 01 2011 - 11:00:07 PDT
