A Parallel Irregular Wavefront Algorithm for Importance Sampling of Probabilistic Networks on GPU

Importance sampling is a widely-used method for probabilistic inference with Bayesian probabilistic networks. Importance sampling is relatively easy to parallelize and parallel GPU implementations yield significant speedups over single-CPU implementations. However, because of physical limitations of GPU memory size and bandwidth, the maximum speedups that can be achieved are bounded by the high data transfer requirements of these algorithms. In this paper, we propose and evaluate a new parallel irregular wave front algorithm for importance sampling of probabilistic networks on GPU. Performance results show that the proposed parallel algorithm achieves greater speedups due to the optimal local memory access compared to simple parallel GPU implementations.