AMBER Archive (2008)

Subject: RE: AMBER: Compile AMBER 9 on TACC Ranger super computer

From: Ross Walker (ross_at_rosswalker.co.uk)
Date: Thu Aug 14 2008 - 11:00:02 CDT


> well. It is a little bit complex to compile Amber9 in some special
> platform.

I'd like to point out that this is not really true here. Sure on something
exotic like a IBM Blue Gene or Cray X1E where you are cross compiling it can
be tricky but Ranger is nothing more than a standard infiniband cluster, all
be it a large one.

Most problems that occur are generally a function of a poorly setup machine
environment, confusion over compiler versions and mpi libraries or a user
trying to tweak with things too much and getting confused. The nice thing
about the NSF supercomputers is that they are generally set up very well and
indeed Ranger is no exception here. There are multiple compilers and mpi
libraries available but all the environment options are well documented. If
you follow their guidelines for selecting mpi libraries, compilers etc then
you get an environment that is correctly set up. Which pgf90 will yield the
Portland Group compiler version you want along with the correct MPI
compilation for this compiler version. Additionally all of the development
libraries etc are all installed. So in fact something like Ranger is
generally much easier to compile and run on than some in house cluster. The
problems occur when people start messing with their environment options
directly rather than reading the documentation for the machine on how to
select different compilers etc.

> > /usr/bin/ld: skipping incompatible
> > ...
> >/usr/bin/ld: cannot find -lmpichf90

> In your case one. the error come from the "libmpichf90" file. Are you sure
> this
> file locate on this direction:/share/home/00654/tg458141/local/mpich2-
> 1.0.7/lib/ ?

No the error the user is seeing here is the 'skipping incompatible',
libmpichf90 does indeed exist, the issue is that this was compiled as a 64
bit binary while the amber pgf90 configure script forces 32 bit
compilation...

> In your case two. it is clear taht C compiler can not be found. I suggest
> you to
> use the icc( intel C compiler) and ifort(intel fortran compiler).

No, one should use the default c compiler on ranger, which is gcc - it works
fine. The issue here is that the default environment has been changed in
some way resulting in the autoconfigure of netcdf (not amber) for some
reason finding that the c compiler is not working. One option is to skip the
-bintraj flag to configure so that netcdf is not built, the other is to
restore the default environment variables so the c compiler works again.
 
> by the way, "-tp p7" or "-tp althon" is decided by the CPU of you
> supermachine.
> Please use "man" command to check the option of pgf90 or ifort to ensure
> which is followed to "-tp"

This has been discussed numerous times on the mailing list before. At the
time of Amber 9's release there were bugs in the pgf90 64 bit compilation
routines that lead to incorrect results being produced and thus failures of
the test cases. This is very very serious. A segfault is one thing but
silently generating incorrect data is very bad. This is why we have the
comprehensive test suite to detect such problems. At the time this problem
with pgf90 was found it was very close to release and we couldn't find a
workaround for the compiler issues other than to turn off 64 bit compilation
altogether. Hence why the -tp p7 was deliberately added for an 'opteron'
build. It forces 32 bit compilation.

The issue is that to build in parallel you have to link against all 32 bit
mpi libraries which generally means that the administrator of the machine
must have built both sets for you, the 64 bit default ones and the 32 bit
ones.

Unfortunately on Ranger only the 64 bit versions are available hence the
linking problems with pgf90. A simple solution is to remove all the flags
that force 32 bit compilation. The latest version of pgf90 on ranger

login3% pgf90 -V

pgf90 7.1-2 64-bit target on x86-64 Linux -tp gh-64
Copyright 1989-2000, The Portland Group, Inc. All Rights Reserved.
Copyright 2000-2007, STMicroelectronics, Inc. All Rights Reserved.

has hopefully fixed the 64 bit compiler problems.

Indeed, I have just tried this out on Ranger without encountering any
problems. Here is what I did - fairly simple really:

tar xvzf amber9.tgz
setenv AMBERHOME ~/amber9
cd amber9
wget http://www.ambermd.org/bugfixes/9.0/bugfix.all
patch -p0 -N -r patch-rejects <bugfix.all

..> login4% which mpif90
..> /opt/apps/pgi7_1/mvapich2/1.0/bin/mpif90
..> login4% which pgf90
..> /opt/apps/pgi/7.1/linux86-64/7.1-2/bin/pgf90

setenv MPI_HOME /opt/apps/pgi7_1/mvapich2/1.0/

./configure -mpich2 -opteron pgf90

edit config.h and remove all instances of -tp p7 and -m32 (otherwise you
can't link against the 64 bit mvapich2)

make parallel

Next we test the parallel implementation...

---- qsub script ----
#!/bin/bash
#$ -V # Inherit the submission environment
#$ -cwd # Start job in submission directory
#$ -N testPMEMD # Job Name
#$ -j y # combine stderr & stdout into stdout
#$ -o $JOB_NAME.o$JOB_ID # Name of the output file (eg. myMPI.oJobID)
#$ -pe 2way 32 # Requests 2 cores/node, 32/16 = 2 nodes total (4
cpu)
#$ -q development # Queue name
#$ -l h_rt=00:45:00 # Run time (hh:mm:ss) - 0.5 hours
 set -x #{echo cmds, use "set echo" in csh}
 setenv AMBRHOME ~/amber9
 cd $AMBERHOME/test
 setenv DO_PARALLEL ibrun
 make test.parallel

-------------------------------------------

Note there are several failures here - mainly in hybrid REMD, LES and PIMD
so stay away from these types of runs with this installation on RANGER!

Next we build PMEMD - this just needs a minor modification to work with
mvapich2 since it only has options for mvapich (basically tries to link to a
non-existance mtl-common otherwise).

cd $AMBERHOME/src/pmemd/
./configure linux64_opteron pgf90 mvapich pubfft nobintraj
..> Please enter name of directory where Infiniband libraries are installed:
..> /opt/apps/pgi7_1/mvapich2/1.0/lib/

edit config.h and change MPI_LIBS line to MPI_LIBS =

change all pgf90's to mpif90

make install

Next we test the PMEMD implementation...

---- qsub script ----
#!/bin/bash
#$ -V # Inherit the submission environment
#$ -cwd # Start job in submission directory
#$ -N testPMEMD # Job Name
#$ -j y # combine stderr & stdout into stdout
#$ -o $JOB_NAME.o$JOB_ID # Name of the output file (eg. myMPI.oJobID)
#$ -pe 2way 32 # Requests 2 cores/node, 32/16 = 2 nodes total (4
cpu)
#$ -q development # Queue name
#$ -l h_rt=00:45:00 # Run time (hh:mm:ss) - 0.5 hours
 set -x #{echo cmds, use "set echo" in csh}
 setenv AMBRHOME ~/amber9
 cd $AMBERHOME/test
 setenv DO_PARALLEL ibrun
 make test.pmemd

-------------------------------------------

These tests all pass without problems.

All done...

Again this may not give optimum performance, you can tweak to your hearts
content to get better performance out but this should at least be reasonable
and will give a working (and tested assuming you check the results from the
test runs) version of AMBER 9.

Have fun....
Ross

/\
\/
|\oss Walker

| Assistant Research Professor |
| San Diego Supercomputer Center |
| Tel: +1 858 822 0854 | EMail:- ross_at_rosswalker.co.uk |
| http://www.rosswalker.co.uk | PGP Key available on request |

Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" (in the *body* of the email)
      to majordomo_at_scripps.edu