Achieving Computationally Scalable Parallelism in a Geostatistical Inversion Algorithm

We present a systematic approach to achieving computationally scalable parallelism in the context of geostatistical inversion. Our experience shows that efficient utilization of hardware requires recursive application of domain decomposition, ranging from multi-process model on a cluster of workstations to multithreading on individual CPU cores. Actual run-times and the degree of realistically achievable scalable parallelism depend on a multitude of factors. The list includes the project area size, the probabilistic model complexity, coarse scale of a stochastic iteration of the Multigrid Monte Carlo scheme, and hardware specifications, to name a few. However, overall, we were able to achieve close to optimal multithreaded scalability (where theoretically possible) on up to 32 CPU cores. We also observe that efficiency of multithreading becomes limited by the memory bus bandwidth at some point.