equilibrium: Vectorizing the refinement step about pycalphad HOT 4 CLOSED

We have pretty good performance with the current sequential solver, with point calculations typically taking less than 5 seconds. The trick is getting more of the steps to be performed sequentially. I'm leaning toward a rewrite of the refinement loop in Cython to make the whole thing multi-threaded.

from pycalphad.

Another possibility, which would complicate Python 2 compatibility: https://gist.github.com/mangecoeur/9540178

from pycalphad.

The new ufuncify-based backend is providing very nice solver performance at minimal compilation cost. If we could reliably parallelize the function constructor for each phase, that would be pretty close to ideal. Writing the refinement loop in Cython could provide even better performance, but I need to perform profiling after the compilation step is cleaned up, because it's difficult to get good numbers on the rest of the code when compilation is such a significant bottleneck.

from pycalphad.

With multi-core processing enabled by dask (essentially just leaning on multiprocessing, with better pickling), I think this is essentially done.

from pycalphad.