RE: [AMBER] Compiling pmemd (Amber 10) with OpenMPI, pathscale, etc.

From: Ross Walker <>
Date: Thu, 22 Apr 2010 16:23:37 -0700

Hi Paul,

A fix is coming for this in a few minutes. It is related to issues with
bugfix.31. Try the attached patch file and let me know if it works.

cd $AMBERHOME/src/pmemd/src/
cp /path_to_patch/patch ./
patch -p0 < patch
cd ../

With regards to building for openMPI etc. The problem is the PMEMD 10 build
scripts all try to guess at the MPI libraries rather than using mpif90 and
mpicc which can leave one in a world of trouble. This has been addressed in
AMBER 11 and it should be much simpler to build things.

For the time being what you did in the config.h file is correct. Basically
empty the MPI_LIBS line and replace all occurrences of the C or Fortran
compilers with mpicc and mpif90 respectively. Note the optimization flags
are hopelessly out of date for PGF90 since I don't think any of the
developers have access to newer versions of the Portland group compilers,
preferring to stick with Intel and/or Gnu. What you have though looks good
to me.

I wouldn't bother too much with FFTW. It makes a marginal difference at best
and tends to just complicate the build process.

All the best

> -----Original Message-----
> From: [] On
> Behalf Of Paul S. Nerenberg
> Sent: Thursday, April 22, 2010 3:58 PM
> To:
> Subject: [AMBER] Compiling pmemd (Amber 10) with OpenMPI, pathscale,
> etc.
> Hello AMBER community,
> Over the last couple of weeks, I've tried (very unsuccessfully!) to
> get pmemd compiled on our group's Linux cluster, which consists of a
> mixed set of Opteron- and Xeon-based nodes. Looking through the list
> archive I've found numerous posts related to pmemd and OpenMPI, and
> implementing the knowledge gained from these posts, I've managed to
> get pmemd to compile without displaying any errors (see [1]). But
> when I run test.pmemd -- or short jobs that I know run fine with
> sander -- immediately get MPI errors (see [2]). (N.B. I have "make
> install" generate a binary called pmemd-test and have updated the
> Makefile in the /test directory to reflect that change...just in case
> anyone thought that was the problem.)
> I have successfully compiled, tested, and run an MPI version of sander
> to my heart's content (see [3] for my config_amber.h), so I don't
> think there are any problems with our pathscale or OpenMPI
> installations. I've also tried using the pubfft option when compiling
> pmemd (instead of fftw), but get the same MPI errors from the compiled
> binary either way. I've posted my config.h file for pmemd (see [4])
> so that folks can take a look at that and make suggestions. (I should
> note that I've also tried leaving the MPI_LIBS field blank and that
> did not remedy things.) Hopefully it's something obvious that I'm
> just missing...
> Many thanks in advance,
> Paul
> [1]
> [2]
> [3]
> [4]
> _______________________________________________
> AMBER mailing list

AMBER mailing list

Received on Thu Apr 22 2010 - 16:30:23 PDT
Custom Search