Parallelisation issues

[gmingas]

I posted this in pymc3-devs/pymc3 regarding the inability to run multi-core sampling with theano ops in my Mac:
pymc-devs#4085

@mikkelbue not sure if your issue is related but take a look at the other issues I mention in there in case you find something that connects.

[gmingas]

It seems the reason parallelisation does not work along with a workaround is here, i.e. when running on Mac you need to import from an external file all the classes and functions that you use to define your likelihood.

[gmingas]

See this: pymc-devs#4085

[gmingas]

We have at least two issues:

• On MacOS the parallelisation does not work on jupyter when Theano ops are used (mentioned above). See this: Parallel sampling with Theano Op fails on MacOS 10.15 pymc-devs/pymc#4085
• On linux sometimes parallel runs are not faster than single thread runs.

[mikkelbue]

I have two parallelisation issues on my Linux (Ubuntu 18.04) machine, really:

1. Like you mentioned, parallel runs take as long as serial ones, but for the Gravity example only. Si I expect it might be something on the NumPy side, since that is the only library that is used in that example.
2. If using FEniCS in the forward model, the Jupyter kernel crashes and immediately restarts as soon as sampling is complete. It does run faster than serial, though, but the results are lost after the kernel restart. I suspect this has something to do with the order of imports, but I am not sure.