Optimizing RDF Stores by Coupling General-purpose Graphics Processing Units and Central Processing Units

From our experience in using RDF stores as a backend for social media streams, we pinpoint three shortcomings of current RDF stores in terms of aggregation speed, constraints checking and largescale reasoning. Parallel algorithms are being proposed to scale reasoning on RDF graphs. However the current efforts focus on the closure computation using High Performance Computing (HPC) and require prematerialization of the entailed triples before loading the generated graph into RDF stores, thus not suitable for continuously changing graphs. We propose a hybrid approach using General-purpose Graphics Processing Units (GPGPU) and Central Processing Units (CPU) in order to optimize three aspects of RDF stores: aggregation, constraints checking, and dynamic materialization.