best parallelisation strategy on python

[raymond.nusbaum at gmail.com]

On Wednesday, April 18, 2018 at 7:16:19 PM UTC-6, simona bellavista wrote:
> I have a code fortran 90 that is parallelised with MPI. I would like to traslate it in python, but I am not sure on the parallelisation strategy and libraries. I work on clusters, with each node with 5GB memory and 12 processors or 24 processors (depending on the cluster I am using). Ideally I would like to split the computation on several nodes.
> 
> Let me explain what this code does: It read ~100GB data, they are divided in hdf5 files of ~25GB each. The code should read the data, go through it and then select a fraction of the data, ~1GB and then some CPU intensive work on it, and repeat this process many times, say 1000 times, then write the results to a single final file.
> 
> I was thinking that the CPU intensive part would be written as a shared object in C.
> 
> Do you have suggestions about which library to use?

I would suggest the Python multiprocessing package. In Python you have to use processes to get full parallelism as there is a single lock on the Python interpreter. The multiprocessing package supports this computing model.