Hi Dawei,
My guess would be that you have not renamed EVERY single file that sander can output. This includes things like mdinfo. Make sure you specify every output file so you do not get two threads trying to write to the same file. This includes profile_mpi written at the end of a run, which is turned off by default in AMBER 10 I believe. Essentially many times when people run on clusters using either NFS or unfortunately even parallel file systems like lustre which really should no better, they let multiple threads write to the same file. The file system then employs locking to make sure each write is atomic and performance can fall off by 1000% or worse. When all writes are done from the same node the locking is handled inside the node's own write cache so you do not see the effect of it.
Hence make sure you specify EVERY output file on the command lines in the group file - all to different files.
Good luck,
Ross
> -----Original Message-----
> From: amber-bounces.ambermd.org [mailto:amber-bounces.ambermd.org] On
> Behalf Of Da-Wei Li
> Sent: Wednesday, November 25, 2009 9:28 AM
> To: AMBER Mailing List
> Subject: Re: [AMBER] Verlet update time in explicit water Replica
> Exchange MD
>
> Hi,
>
> Just another update. If I re-arrange the MPI machine file to let the
> master processors of each group to be on the same node, everything
> will be fine. I have no idea why this happens.
>
> have a nice day.
>
> dawei
>
> _______________________________________________
> 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 Wed Nov 25 2009 - 13:00:02 PST
