Test the impact of keeping allmodels into shared-memory (multiproc branch)

[pcubillos]

If we keep it in memory we can enable the savemodel argument. Shared memory is the best option I can think off (certainly better than sending data through pipes).

[pcubillos]

Nate had a much better idea: Keep the model arrays in local memory for each chain, and return the final array to mcmc, just once, at the end of the MCMC run.