- Operating System: “Red Hat Enterprise Linux Server 7.2 (Maipo)”
- brms Version: 2.9.0
- rstan version: 2.19.2
Recently I have been working on what are essentially generalized additive models (GAMs), initially using mgcv. However, for both computational and principled reasons I would like to move to a Bayesian framework. I am still pretty new to both GAMs and Bayesian inference, so specifying a GAM directly in rstan is beyond me, at least for now. I have therefore been using the amazing brms package to do things like this (reduced version):
Family: exponential Links: mu = log Formula: eeg06 ~ s(relright, k = 20) + s(relleft, k = 20) Data: subset(eeg, select = c(deps[d], "relright", "relle (Number of observations: 5266)
However, it has quickly become clear that running these analyses on my office pc is not feasible: even a considerably reduced model takes about 4 days to run. I do have access to a supercluster, but my initial attempts don’t do much better.
From what I can find, the newest version of RStan supports MPI parallelization. My question is: does brms make use of this automagically (i.e., I cannot achieve any further performance gains)? And if not, can I set up brms for this? Or will I need to invoke rstan directly/‘manually’ to use the MPI features?
I would be most grateful for any guidance you can provide on this.