Massively parallel MCMC for Bayesian hierarchical models