scalapy is a wrapping of Scalapack such that it can be called by Python in
a friendly manner.
Operations are performed on DistributedMatrix objects which can be easily
created whilst hiding all the nasty details of block cyclic distribution.
This package depends upon two python packages numpy and mpi4py. It is
written largely in pure Python, but some parts require Cython and
f2py. Obviously it also requires an MPI distribution (OpenMPI and
IntelMPI supported out the box), and a Scalapack installation (both Intel
MKL and NETLIB are currently supported).
To build just use the standard setup.py script:
$ python setup.py install
It will attempt to probe you current environment to determine which MPI
distribution, and ScaLAPACK installation to use. As this isn't completely
robust, you can edit setup.py manually specify what to use.
Limited, but improving, documentation is available here.
Some of the features, especially distribution of matrice from global arrays
and files, make heavy use of advanced features of MPI, such as derived
datatypes and MPI-IO. Unfortunately many MPI distributions contain critical
bugs in these components (mostly due to ROMIO), which means these will
fail in some common circumstances.
However, recent versions of OpenMPI contain a new implementation of MPI-IO (called OMPIO) which seems to be issue free. This means that for full, and successful usage you should try and use OpenMPI 1.8.2 or greater. Additionally, you may need to force it to use OMPIO rather than ROMIO. This can be done by calling with:
$ mpirun -mca io ompio ...
or by setting the environment variable:
$ export OMPI_MCA_io=ompio