[Paper Review] Fast Processing of SPARQL Queries on RDF Quadruples

Vasil Slavov, Anas Katib+3 more|arXiv (Cornell University)|Jun 3, 2015

Semantic Web and Ontologies21 references4 citations

TL;DR

This paper proposes RIQ, a novel indexing and query processing approach for large-scale RDF datasets with quadruples (quads). By grouping similar RDF graphs and using a filtering index based on Bloom Filters, RIQ reduces query processing time through a decrease-and-conquer strategy, outperforming RDF-3X and Jena TDB on complex SPARQL queries over datasets with ~1.4 billion quads.

ABSTRACT

In this paper, we propose a new approach for fast processing of SPARQL queries on large RDF datasets containing RDF quadruples (or quads). Our approach called RIQ employs a decrease-and-conquer strategy: Rather than indexing the entire RDF dataset, RIQ identifies groups of similar RDF graphs and indexes each group separately. During query processing, RIQ uses a novel filtering index to first identify candidate groups that may contain matches for the query. On these candidates, it executes optimized queries using a conventional SPARQL processor to produce the final results. Our initial performance evaluation results are promising: Using a synthetic and a real dataset, each containing about 1.4 billion quads, we show that RIQ outperforms RDF-3X and Jena TDB on a variety of SPARQL queries.

Motivation & Objective

• To address the performance bottleneck in processing complex SPARQL queries on large RDF datasets containing billions of RDF quads.
• To overcome the limitations of existing approaches that rely on full dataset indexing and expensive join operations for complex graph patterns.
• To reduce query processing time by identifying and focusing on candidate graph groups that may contain matches, rather than scanning the entire dataset.
• To design a compact, efficient filtering index using Bloom Filters and Counting Bloom Filters to support fast candidate group identification.

Proposed method

• RIQ represents RDF graphs and SPARQL query patterns using a vector-based similarity model to group semantically similar graphs together.
• It constructs a filtering index using Bloom Filters and Counting Bloom Filters to compactly represent graph group signatures, enabling fast candidate filtering.
• During query processing, RIQ uses the filtering index to identify only the candidate graph groups that may contain matches for the query pattern.
• It then delegates final query execution to a conventional SPARQL processor, which processes only the candidate groups, reducing I/O and computation.
• The approach leverages a decrease-and-conquer strategy: instead of processing all data, it narrows the search space to relevant graph groups.
• The filtering index is built with a 5% false positive rate to balance compactness and accuracy in candidate selection.

Experimental results

Research questions

• RQ1 Can a filtering index based on Bloom Filters effectively reduce the search space for SPARQL queries on large RDF quad datasets?
• RQ2 Does grouping similar RDF graphs and processing only candidate groups lead to significant performance gains over full-scan or join-based approaches?
• RQ3 How does RIQ’s performance compare to RDF-3X and Jena TDB on complex SPARQL queries with large, cyclic graph patterns?
• RQ4 Can RIQ maintain efficiency on both large and small SPARQL queries, including those with optional and union patterns?
• RQ5 What is the trade-off between index size and query performance in RIQ’s graph grouping and filtering strategy?

Key findings

• On the LUBM dataset with ~1.4 billion quads, RIQ achieved a geometric mean query processing time of 144.09 seconds, outperforming RDF-3X (375.98 s) and Jena TDB (448.65 s) in cold cache settings.
• For BTC 2012, RIQ’s geometric mean time was 35.45 seconds, significantly lower than RDF-3X (372 s) and Jena TDB (168.22 s), demonstrating superior performance on real-world data.
• For complex queries with large, cyclic graph patterns (L1-L3, B1-B2), RIQ was substantially faster than both RDF-3X and Jena TDB, with query times up to 77,000x faster in some cases.
• The filtering index in RIQ was only 8.5 GB for LUBM and 16 GB for BTC 2012, while RDF-3X and Jena TDB required 77–121 GB and 87–110 GB respectively, showing a major reduction in index size.
• For small queries (L4-L12, B3-B7), RIQ was faster than both baselines in six out of nine LUBM queries and three out of five BTC 2012 queries in cold cache, though RDF-3X led in warm cache for some small queries.
• RIQ identified only 22 candidate groups at most for complex queries, drastically reducing the number of graphs that needed full processing, confirming the effectiveness of its filtering strategy.