state of MPI support

[Felixrccs]

I started to try out your code and adapt it to my systeme/problem, and it is set up with MPI support in mind. So first how "ready" is the MPI setup and how is it supposed to work?

Is it suppose to start multiple Nested Sampling simulations with a share of the total walkers each?

Or is it supposed to do start one simulation and cull multiple walkers each iteration?

[bernstei]

It's definitely ready - I only use it that way. But, it's neither culling multiple walkers nor doing independent NS simulations with a share of the walkers.

The way parallelism is done is that the walkers are split up among the MPI tasks, and each MPI task walks one walker (at random) for N_steps / N_tasks. This leads to an expected number of steps before culling of N_steps, with the caution that this is now a stochastic average, not an exact walk length. Then the top walker is culled, a random walker is duplicated, and the process repeats. This limits the degree parallelism to N_tasks such that N_steps / N_tasks isn't too small - I typically don't go below around 20 or so.

Culling multiple walkers was available in pymatnest, but was dropped from pymatnext because the message passing communication pattern was much more complex and multiple walker culling never seemed worth it (I recall both increased noise and likelihood of minima extinction) for the sort of test system we benchmarked. It should be possible to add it back, given sufficient motivation.

[bernstei]

Technically, this is written in the README, but admittedly not in an at all explanatory way.