[Pypar-developers] Successful compilation on x86_64 cluster

[Craig F. <cf...@ie...>]

Here' s a short summary of what I did to compile pypar and install it in my home directory on our Intel Xeon 64 bit cluster with Intel compilers and MPI:

Compile and install pypar:
python setup.py config --compiler=intel install --home=~

The mpiext library compiles and installs without error, but it can't be loaded by Python.  I compiled the binary manually with the following command:

icc -fPIC -i-dynamic -D_EM64T_ -D_SMP_ -DUSE_HEADER_CACHING -DONE_SIDED -DMPID_USE_SEQUENCE_NUMBERS -D_SHMEM_COLL_ -DRDMA_CM -I/usr/include/python2.4 -I/home/cfinch/lib64/python/numpy/core/include -I/usr/include -O2 -I/usr/mpi/intel/mvapich2-0.9.8-15/include -L/usr/mpi/intel/mvapich2-0.9.8-15/lib -Wl,-rpath -Wl,/usr/mpi/intel/mvapich2-0.9.8-15/lib -lmpich -L/usr/lib64 -lrdmacm -libverbs -libumad -lpthread -lrt -I/usr/include/python2.4 -c mpiext.c -o mpiext.o

I then linked it using mpicc:
mpicc -shared mpiext.o -o mpiext.so -mcmodel=medium

And copied it to 
cp mpiext.so ~/lib64/python/pypar/

I don't claim to understand the options I had to use to get icc to compile the extension correctly.  I basically looked at the files setup.py and compile.py, and kludged together options until something worked.  It would be great if someone could explain how this process works and include it in the docs for future reference.

I do have one question.  When I submit a pypar job using mpirun, python can't seem to find the Python packages (numpy, pypar, etc). that are installed in my home directory (~/lib64/python).  I verified that ~/lib64/python is in the Python path for the Python processes that are started by mpirun, but they still can't find my packages.  Any suggestions how to make them available?  I don't think our cluster admin has any experience with parallel Python.

   Craig Finch


__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 
http://mail.yahoo.com