Experimental framework for searching large RDF on GPUs based on key-value storage

Abstract

Resource Description Framework (RDF) is commonly used for the semantic web query. During this decade, due to big data processing, the large numbers of RDF triples are crawled. The triples usually stored distributed on the clouds storage or the large clusters. To search for the query answer, it is usually difficult to handle the search across platforms. Also, the search takes a long executed time. Thus, the data representation and platform are important to speedup the search and handle the heterogeneousness. In this paper, we present the experimental framework which can be used to handle the search of RDF data in GPU clusters. Our framework uses the Java platform to manipulate the semantic query while using JCuda1 to perform the GPU processing. Apache Cassandra storage, known as CumulusRDF, is used to store key-values for searching. In the experiments, DBpedia and Freebase dataset are extracted and manipulated. The triple structures are transformed and loaded into Apache Cassandra storage as CumulusRDF's flat layout. The subject-predicate-object keys are kept in CQL caching. There are about 3-hundred-million tags that can be handled with in one machine, which can reduce time, with an inexpensive cost. We shape the data grid from row-major-ordering of Java, to GPU thread grid of CUDA, retrieved keys to join for finding the correspondence of the RDF graph.