Scalable Bayesian Inference of Large Simulations via Asynchronous Prefetching Multilevel Delayed Acceptance

Bayesian inference enables greater scientific insight into simulation models, determining model parameters and meaningful confidence regions from observed data. With hierarchical methods like Multilevel Delayed Acceptance (MLDA) drastically reducing compute cost, sampling Bayesian posteriors for computationally intensive models becomes increasingly feasible. Pushing MLDA towards the strong scaling regime (i.e. high compute resources, short time-to-solution) remains a challenge: Even though MLDA only requires a moderate number of high-accuracy simulation runs, it inherits the sequential chain structure and need for chain burn-in from Markov chain Monte Carlo (MCMC).We present fully asynchronous parallel prefetching for MLDA, adding an axis of scalability complementary to forward model parallelization and parallel chains. A thorough scaling analysis demonstrates that prefetching is advantageous in strong scaling scenarios. We investigate the behavior of prefetching MLDA in small-scale test problems. A large-scale geophysics application, namely parameter identification for non-linear earthquake modelling, highlights interaction with coarse-level quality and model scalability.