Possibility of computing blackbox likelihood using MPI


I am relatively new to Pymc3 and have used it to calibrate my model along with Theano for my likelihood computation. Since for my model (non-algebraic) evaluation (likelihood) takes the most amount of time, I was hoping to use mpi4py to parallelize the likelihood evaluation. In other words my likelihood has different experimental conditions (case1, case2, case3,…case100) and I want to compute each of these cases in parallel. Since there is also a parallelism for pymc3 (nprocs) I was wondering this extra level of parallelism would be a problem or not ?

Did anyone explore this avenue ?


A follow up, can I specify pymc3 to use particular thread and processes in my machine ? similar to mpirun -hosts ?