New article about MCMC and parallel computation
Christian P. Robert, Murray Smith and I have just finished an article (arxiv link) on an improved Independent Metropolis-Hastings (IMH) algorithm that can highly benefit from parallel processing units (see this previous post for an introduction on parallel computation).
Our method named “block Independent Metropolis-Hastings” allows significant decrease of the variance of estimators based on the IMH, while being cost-free in a parallel processing context (and still cheap in a single core context, since you don’t do any additional target density evaluations compared to the standard IMH algorithm).
library(MASS) data(Pima.te) ?Pima.te
Update2: I’ve put the code online here (you need Python, numpy, scipy.weave for the computation part, plus R and rpy2 for the plots).