smpi

Simple MPI for Python allows MPI parallelisation using python decorators.

@smpi.collect(smpi.dist_type.gather)
@smpi.distribute(smpi.dist_type.local, smpi.dist_type.local, smpi.dist_type.scatter)
def calculate_mi(X, Y, features):
    MI = numpy.full([len(features)], numpy.nan, dtype=numpy.float)
    for i,X_i in enumerate(features):
        MI[i] = pymit.I(X[:, X_i], Y , bins=[bins, 2])
    return [MI]

The example uses previously distributed numpy Arrays X and Y and scatters a 1D numpy array features. Than some computation is done, and the result is gathered and returned. Instead of explicitly distributing the data, the user only specifies the operation that needs to be done for each parameter:

• X and Y are already in memory, so smpi.dist_type.local is specified
• Each rank is supposed to compute some features, so it is scattered (i.e. using smpi.dist_type.scatter)
• The result for each feature out of features needs to be collected so gather is used (i.e. using smpi.dist_type.gather)
• Obviously the work needs to be Distributed between the ranks, so smpi.distribute is invoked, and finally collected, so smpi.collect is used.

A more detailed example can be found in hcmi_smpi.py. Details about Mutual Information and Feature Selection (e.g. pymit) can be found at https://github.com/tud-zih-energy/pymit .

Usage:

For smpi itself, you only need mpi4py. To try the example, you need the following:

• Get Madelon data from: http://clopinet.com/isabelle/Projects/NIPS2003/MADELON.zip
• place the data in a MADELON folder
• Get and install pymit https://github.com/tud-zih-energy/pymit in your env.
• Get mpi4py.

After cloning the repo run:

pip install mpi4py
pip install git+https://github.com/tud-zih-energy/pymit.git
wget http://clopinet.com/isabelle/Projects/NIPS2003/MADELON.zip
unzip MADELON.zip
OMP_NUM_THREADS=1 mpirun -n 8 python hcmi_smpi.py ./MADELON/MADELON/ 20

License

If the LGPL is not good License for you, please feel free to write to [email protected]