Exploiting data parallelism and population Monte Carlo on massively-parallel architectures for geoacoustic inversion

Bayesian inference algorithms in geoacoustic inversion have high computational requirements on multiple computational scales. Predicting (modeling) data to match observations represents fine-grained computations which often cannot be implemented efficiently on CPU clusters since high latency and communication overhead outweigh parallelization gains. However, GPUs, which operate efficiently on 100,000s of parallel treads with low latency and high bandwidth, can provide signficant performance gains. Bayesian sampling schemes are generally coarse-grained, and can be implemented efficiently in parallel on multi-core/cluster architectures. For example, population Monte Carlo methods simulate many Markov chains in parallel, with chains running independently between interactions (at predefined intervals) which exchange information throughout the population, substantially increasing sampling efficiency. This paper combines fine- and coarse-grained parallelization to profoundly improve the efficiency of geoacoustic inversion of seabed reflection data. Spherical-wave reflection-coefficient predictions, which require solving the Sommerfeld integral for a large number of grazing angles and frequencies, constitute fine-grained, data-parallel computations which are implemented efficiently on a GPU. Sampling is based on population Monte Carlo simulation with chain interactions as exchange and crossover moves. The algorithm is applied to data from the Malta Plateau to study the frequency dependence of sound velocity and attenuation in marine sediments.