Streaming Partitioning of RDF Graphs for Datalog Reasoning

Temitope Ajileye, Boris Motik, Ian Horrocks

Keywords: datalog, materialisation, streaming, partitioning, distributed, reasoning, RDF, knowledge bases, graphs

Abstract:

A cluster of servers is often used to reason over RDF graphs whose size exceeds
the capacity of a single server. While many distributed approaches to reasoning
have been proposed, the problem of data partitioning has received little
attention thus far. In practice, data is usually partitioned by a variant of
hashing, which is very simple, but it does not pay attention to data locality.
Locality-aware partitioning approaches have been considered, but they usually
process the entire dataset on a single server. In this paper, we present two
new RDF partitioning strategies. Both are inspired by recent \emph{streaming}
graph partitioning algorithms \cite{hdrf:2015,2ps}, which partition a graph
while keeping only a small subset of the graph in memory. We have evaluated our
approaches empirically against hash and min-cut partitioning. Our results
suggest that our approaches can significantly improve reasoning performance,
but without unrealistic demands on the memory of the servers used for
partitioning.

Subtrack: Ontologies and Reasoning

First Author Is Student: Yes