Posted by Serguei Patchkovskii on June 01, 2009 at 23:00:01:
In Reply to: Re: Weird compilation problems with GA in MOLCAS 7.2 posted by Valera Veryazov on May 30, 2009 at 08:20:17:
: well.. all questions about GA installation should be addressed to GA mailing list.
: I (personally) would prefer to exclude GA from the Molcas distribution, and have it as prerequisite library.
I am not so sure. GA is so fragile that I essintially gave up trying to build it as a system-wide support library: I always end up building an application-specific copy of GA for each application which uses it. At least in the setup you have now, there is a reasonable chance that GA will be built with options compatible with the rest of the code.
As far as my initial problem was concerned, I've found the solution:
MPI-dependent parts of GA appears to be incompatible with GCC 4.3 (which is installed by default with openSuSE 11). Building it with gcc 4.2.4 produces a working MPI executable. Further, GA appears not to work with non-TCP transports in OpenMPI 1.2.8. In order to use Infiniband, I had to build with OpenMPI 1.3.2. This lets me use OpenIB
transport in ARMCI - however, GA still won't let me use more than one process per node. Finally, if I use GA 4.1.1, I can use MPI-2 transport and get my accellerated interconnect and can use all cores in each node.
So my GA library and OpenMPI were "too old", while my compiler was "too new". Argh.
If anybody cares, the actual options for building GA 4.1.1 were:
gmake TARGET=LINUX64 FC=ifort FOPT="-O3 -xP -mp" \
FLDOPT="-Wl,-rpath,/usr/local/openmpi/1.3.2/intel/lib,-rpath,/usr/local/gcc/4.2.4/lib64/,-rpath,/usr/local/intel/icte/3.1.1/cce/10.1.015/lib/" \
CLDOPT="-Wl,-rpath,/usr/local/openmpi/1.3.2/intel/lib,-rpath,/usr/local/gcc/4.2.4/lib64/" \
ARMCI_NETWORK=MPI-SPAWN USE_MPI=yes LARGE_FILES=yes \
MPI_INCLUDE=/usr/local/openmpi/1.3.2/intel/include/ MPI_LIB=/usr/local/openmpi/1.3.2/intel/lib/
Have fun,
Serguei