In this paper, we consider the problem of materializing the complete finite RDFS closure in a scalable manner; this includes those parts of the RDFS closure that are often ignored such as literal generalization and container membership properties. We point out characteristics of RDFS that allow us to derive an embarrassingly parallel algorithm for producing said closure, and we evaluate our C/MPI implementation of the algorithm on a cluster with 128 cores using different-size subsets of the LUBM 10,000-university data set. We show that the time to produce inferences scales linearly with the number of processes, evaluating this behavior on up to hundreds of millions of triples. We also show the number of inferences produced for different subsets of LUBM10k. To the best of our knowledge, our work is the first to provide RDFS inferencing on such large data sets in such low times. Finally, we discuss future work in terms of promising applications of this approach including OWL2RL rules, MapReduce implementations, and massive scaling on supercomputers.
展开▼
