Out-of-Core GPU-Accelerated Causal Structure Learning

Learning the causal structures in high-dimensional datasets enables deriving advanced insights from observational data. For example, the construction of gene regulatory networks inferred from gene expression data supports solving biological and biomedical problems, such as, in drug design or diagnostics. With the adoption of Graphics Processing Units (GPUs) the runtime of constraint-based causal structure learning algorithms on multivariate normal distributed data is significantly reduced. For extremely high-dimensional datasets, e.g., provided by The Cancer Genome Atlas (TCGA), state-of-the-art GPU-accelerated algorithms hit the device memory limit of single GPUs and consequently, execution fails. In order to overcome this limitation, we propose an out-of-core algorithm for GPU-accelerated constraint-based causal structure learning on multivariate normal distributed data. We experimentally validate the scalability of our algorithm, beyond GPU device memory capacities and compare our implementation to a baseline using Unified Memory (UM). In recent GPU generations, UM overcomes the device memory limit, by utilizing the GPU page migration engine. On a real-world gene expression dataset from the TCGA, our approach outperforms the baseline by a factor of 95 and is faster than a parallel Central Processing Unit (CPU)-based version by a factor of 236.