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.
scalapy
supports both Python 2 and 3 (2.7, 3.2 or later).
The package depends upon two python packages numpy
and mpi4py
. It is
written largely in pure Python, but some parts require Cython
and f2py
.
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 matrices 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