Reverse Partitioning for SPARQL Queries: Principles and Performance Analysis

Abstract

RDF and SPARQL have been widely adopted for modeling and querying Web objects as facts in the Semantic Web. The amount of data stored in RDF format has grown significantly pushing RDF processing systems to implement efficient query processing techniques in parallel and distributed architectures. In such environments, the data partitioning is a pre-condition for query performance. Traditionally, the graph-based RDF systems store the data using adjacency lists formed by a vertex and its outgoing edges. Nevertheless, for a certain type of queries, considering entities and their ongoing edges may speed up their execution. This point motivates us to present a new partitioning technique (called reverse partitioning) dedicated to graph-based triple stores that is complementary to traditional ones. In this paper, we first detail its main principles by illustrating its functioning. Secondly, the best classes of queries for which reverse partitioning gives better performance are discussed. Finally, we report on intensive experiments using large RDF datasets that show significant performance improvements for certain queries in a graph-based triple store and in a relational-based system.