Time-consuming simulations run on a massive scale to understand earthquakes better and prepare for possibly devastating scenarios. We cannot locate earthquake sources directly but only determine them with uncertainty from receiver data. We attempt to speed up a Markov chain Monte-Carlo (MCMC) uncertainty quantification of earthquake sources. For this purpose, we explore fusing simulations of the model evaluation and a parallel MCMC algorithm called Generalized Metropolis-Hastings.
We perform simulations on a high-performance Linux cluster to evaluate the runtime performance and convergence rate properties for three design approaches. We confirm that fused simulations can cut the simulation time by half on the same number of hardware nodes. Also, if we keep the ratio of nodes per fused simulation constant, we find good, strong scaling behavior and calculate the parallel region as around 96% of the program. We note that we performed many runs of small problem sizes rather than one extensive simulation. Therefore, 16 fused simulations on 64 nodes take 58% more time per simulation than 16 parallel non-fused simulations on four nodes each.