[molpro-user] Molpro 2008.1 and mvapich

[Jyh-Shyong]

Dear Molpro Users,

I have tried to install Molpro 2008.01 on our cluster, the serial 
version is successful, but the
parallel version with mvapich ( MPICH on infiniband) failed.  I wonder 
if any one has successfully
install Molpro 2008.01 on such environment?

Here is my procedure:

1. Build ga-4.0.8 based on mvpaich
export TARGET=LINUX64
export MPI_INCLUDE=/opt/vltmpi/OPENIB/mpi.icc.rsh/include
export MPI_LIB=/opt/vltmpi/OPENIB/mpi.icc.rsh/lib
export USE_MPI=yes
export BLAS_I8=yes
export BLAS_LIB="-L/opt/intel/cmkl/lib/em64t -lmkl_em64t -lguide 
-lpthread -ldl -lsysfs"
export LIBMPI="-L/opt/vltmpi/OPENIB/mpi.icc.rsh/lib -lmpich -lfmpich 
-L/usr/local/ofed/lib64 -libverbs"
export LARGE_FILES=TRUE
export ARMCI_NETWORK=OPENIB
export IB_INCLUDE=/usr/local/ofed/include
export IB_LIB=/usr/local/ofed/lib64
export PATH=${PATH}:/opt/vltmpi/OPENIB/mpi.icc.rsh/bin
export GA_C_CORE=yes
gmake TARGET=LINUX64 FC="ifort -no-ipo" CC="icc -no-ipo" CXX=icpc


2. Build Molpro 2008.01

./configure -mpp -mppbase /package/chem/lib/ga408 -ifort -icc
make install

3. I modified the definition of LAUNCHER in file bin/molpro as
LAUNCHER="/opt/vltmpi/OPENIB/mpi/bin/mpirun_ssh -hostfile %h -np %n %x"

and add an IF structure for loadleveler:
...
if   [ x"$PBS_NODEFILE" != x ] ; then
 nodefile="$PBS_NODEFILE" # PBS
elif [ x"$QSUB_NODEINF" != x ] ; then
 nodefile="$QSUB_NODEINF" # NQS
elif [ x"$PE_HOSTFILE"  != x ] ; then
 nodefile="$PE_HOSTFILE"  # SunGridEngine
elif [ x"$LOADL_HOSTFILE"  != x ] ; then
 nodefile="$LOADL_HOSTFILE"  # Loadleveler
else
 nodefile=""
fi
...

4. Submit a test job to loadleveler with a job script which contains a 
command
/package/chem/molpro2k8.1/binpp/molpro -n $NCPU -d${JOBDIR}  -o 
${JOB}.out ${JOB}.com


The error message in file $JOB.err is
[0] Abort: [iris333:0] Got completion with error, 
Errorcode=IBV_WC_LOC_LEN_ERR, vendor code=69, dest host rank = 4
 at line 410 in file viacheck.c
[iris333.sro.nchc.org.tw : 0 ] call MPI_Abort remote host is 
[iris335.sro.nchc.org.tw : 4 ].
Cleaning up all processes ...
Some rank on 'iris335.sro.nchc.org.tw' exited without finalize.
done.


Any suggestion?

Jyh-Shyong Ho, Ph.D.
Research Scientist
National Center for High Performance Computing
Hsinchu, Taiwan, ROC