These days I have started my investigations on the actual implementation of the MPI parallelisation in fem-fenics. I found out some information that I will point out here, together with my goals for the next weeks.
First of all, apparently MPI can be used without user intervention on the serial code. This is a feature that DOLFIN boasts, but I would expect it not to pass on to fem-fenics, at least not without some effort on the implementation side. Furthermore, DOLFIN offers also a wrapper for MPI functionalities, thus probably it can be helpful in managing data transfers among threads in C++ code.
An issue that will need to be addressed is making ufl.m robust to parallel execution, since its serial implementation leads to all workers trying to open the same file, thus leading to an error that stops computation. Anyway, even if they could all open the file and write to it, this would entail that lines are copied in random order or more than once, so it must be fixed.
In the end, it seems that the partitioning procedure produces matrices that are not slices of the one assembled in serial execution. Due to this fact, I must go deep in the algorithm to find out the proper way to merge the pieces and obtain the complete matrix, which will be stored as octave_value to allow for further computation using Octave's features.