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