AMBER Archive (2009)

Subject: Re: [AMBER] pmemd compile error....

From: Robert Duke (rduke_at_email.unc.edu)
Date: Thu Mar 19 2009 - 10:24:29 CDT


Hmmm. Insufficient information. For some reason, mpi initialization seems
to be failing in the 7th task. This may be a myriad of things, but I would
tend to suspect some sort of library or compilation problems. If I were
dealing with this personally, I would do the following:
1) Take a known simple testcase - say the JACS benchmark.
2) Compile the uniprocessor version of pmemd and run it on this machine -
any problems?
3) If not, compiler sander or mpi test code on this beast - does sander run
the problem okay? There could be all kinds of "interesting" library and mpi
issues.
4) If all that works, recompile parallel pmemd, but set F90_OPT_DFLT to
$(F90_OPT_DBG). This may give you a coredump from which you can get more
info.

It looks to me like you are not even getting through mpi initialization
here, which basically says it is not a pmemd problem, per se, but something
about how you are compiling/linking in this environment. I really don't
know though, given the limited info.

Oh one thing. You say you "recompiled with 64bit compiler because the HP
supercomputer has more than 1 GB stack limit". Okay, BAD move. You
absolutely don't want to be matching 32 and 64 bit code unless the mixing is
supported by the libraries, and generally there are libraries that will be
used for a 32 or 64 bit executable. You simply can't get into memory
problems with pmemd on a 32 bit executable with more than say 4 processors,
and it costs typically at least 5% to go into 64 bit mode due to the
expansion of code/data associated with the use of 64 bit instead of 32 bit
pointers. So if you dinked with the build params, undink. The pmemd code
works on some systems with a 64 bit build (like ibm aix) but it is slower,
and tricky to get linked right, and can die in various nasty hard to
diagnose ways. So "64-bit good, 32-bit bad, is highly faulty logic". It is
only good to use 64 bits if you have massive amounts of data, and you don't.
Attached is a graphic showing resident memory requirements of pmemd (on 4
processors), and the stack memory is a small fraction of the total memory
used. And note, you can't run problems larger than 999,999 atoms with
pmemd, but that is a file format problem, not a memory usage problem. Now
you CAN get into memory problems with pmemd, but it requires 1) running on 1
or a very few processors, 2) using a really large problem - multiple 100's
of thousands of atoms, AND 3) setting the cutoff to something like half the
boxsize - then the pairlist will be just huge, and you can blow up on that.
Be sure to use an 8 or 9 angstrom cutoff in pmemd running pme - very little
in the way of good reasons to do otherwise.

Hope this helps - Bob Duke
(getting any new system up that has not been previously configured can be
sort of a pain - sorry)

----- Original Message -----
From: "Giorgos Lamprinidis" <lamprinidis_at_pharm.uoa.gr>
To: "AMBER Mailing List" <amber_at_ambermd.org>
Sent: Thursday, March 19, 2009 11:06 AM
Subject: [AMBER] pmemd compile error....

Duke,

My colleague Giannis Georgiadis re-compiled pmemd with 64bit compiler
because our HP supercomputer has more than 1GB stack limit for 64bit
applications. We shall increase the stack limit for the 32bit version to
400MB (this is the maximum) when he reboot the computer.

The warning messages disappeared but the Terminated Abnormally errors are
still there without any other messages..

...................

STOP PMEMD Terminated Abnormally!

STOP PMEMD Terminated Abnormally!

STOP PMEMD Terminated Abnormally!

STOP PMEMD Terminated Abnormally!

STOP PMEMD Terminated Abnormally!

STOP PMEMD Terminated Abnormally!

STOP PMEMD Terminated Abnormally!

STOP PMEMD Terminated Abnormally!

MPI Application rank 6 exited before MPI_Init() with status 0

Failed sending message: (Unable to connect to socket (Connection refused)).

....................

Any suggestions plz?

Dr George Lamprinidis
_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber

_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber