Experiences with Virtuoso Cluster RDF Column Store

Abstract

Virtuoso Column Store [185] introduces vectorized execution into the Virtuoso DBMS. Additionally, its scale-out version, that allows running the system on a cluster, has been significantly redesigned. This article discusses advances in scale-out support in Virtuoso and analyzes this on the Berlin SPARQL Benchmark (BSBM) [101]. To demonstrate the features of Virtuoso Cluster RDF Column Store, we first present micro-benchmarks on a small 2node cluster with 10 billion triples. In the full evaluation we show one can now scale-out to a BSBM database of 150 billion triples. The latter experiment is a 750 times increase over the previous largest BSBM report, and for the first time includes both its Explore and Business Intelligence workloads. The storage scheme used by Virtuoso for storing RDF Subject-PropertyObject triples pertaining to a Graph (hence we have quads, not triples) consists of five indexes: PSOG, POSG, SP, OP, GS. To be precise, PSOG is a B-tree with key (P,S,O,G), where P is a number identifying a property, S a subject, O an object and G the graph. Additionally, there is a B-tree holding URIs and a B-tree holding string literals, both of them used to encode string(-URI)s into numerical identifiers. Users may alter the indexing scheme of Virtuoso but this almost never happens. The three last indexes (SP, OP, GS) are projections of the first two covering indexes, containing only the unique combinations – hence these are much smaller. We note that Virtuoso Column Store Edition (V7) departs from the previous Virtuoso editions (V6) in that