Optimising memory management for Belief Propagation in Junction Trees using GPGPUs

Belief Propagation (BP) in Junction Trees (JT) is one of the most popular approaches to compute posteriors in Bayesian Networks (BN). Such approach has significant computational requirements that can be addressed by using highly parallel architectures (i.e., General Purpose Graphic Processing Units) to parallelise the message update phases of BP. In this paper, we propose a novel approach to parallelise BP with GPGPUs, which focuses on optimising the memory layout of the BN tables so to achieve better performance in terms of increased speedup, reduced data transfers between the host and the GPGPU, and scalability. Our empirical comparison with the state of the art approach on standard datasets confirms significant improvements in speedups (up to +594%), and scalability (as our method can operate on networks whose potential tables exceed the global memory of the GPGPU).