Combining Graph Exploration and Fragmentation for Scalable RDF Query Processing

The flexibility offered by the Resource Description Framework (RDF) has led it to become a very popular standard for representing data with an undefined or variable schema using the concept of triples. Its success has resulted in many large scale multidisciplinary datasets, that have prompted the development of efficient RDF processing systems. Current approaches can be distinguished into two groups: the first, adopting the relational model storing the triples in tables, and the second creating data structures that model RDF data as a graph. The strategies of the first group are more easily scalable since they apply optimization strategies from the relational model like indexing and fragmentation. However, these approaches suffer many overheads when dealing with complex queries (e.g. compounded SPARQL graphs involving filters) persistent in existing applications. On the other hand, graph-based systems that use more complex data structures fail to efficiently manage the main memory and are not scalable in computer hardware with limited resources. In this paper, we propose a novel approach to perform queries (Basic Graph Patterns, Wildcards, Aggregations and Sorting) on RDF data. We propose to combine both RDF graph exploration with physical fragmentation of triples. In this work, we describe our graph-based storage and query evaluation models. Then, we detail the architecture of our system and we largely explain the strategy, based in the Volcano execution model, used to manage the main memory at query runtime. We conducted extensive experiments on synthetic and real datasets to evaluate the efficiency of our proposal. We compared our performance with a relational-based (Virtuoso), a graph-based (gStore) and an intensive-indexing (RDF-3X) approach. According to our evaluation, our system offers the best compromise between efficient query processing and scalability.