I did some performance gathering of JRDF's implementation of SPARQL using some of the FOAF data from Mindswap.
Average for Query 1, 100,000 triples:
* Jena (ARQ 0.9.2) - 14685 ms
* JRDF and JRDF using Tuple Subsumption - 3652 ms
There is only one UNION implementation in JRDF.
Average for Query 2, 100,000 triples:
* Jena (ARQ 0.9.2) - 22872 ms
* JRDF - 6615 ms
* JRDF using Tuple Subsumption - 3750 ms
Average for Query 3, 100,000 triples:
* Jena (ARQ 0.9.2) - 15306 ms
* JRDF - 8019 ms
* JRDF using Tuple Subsumption - 4780 ms
The point is not that it was faster that Jena (although yay!), it's that the relational optimisations had a positive effect on querying speed. The current downloadable version of the JRDF SPARQL GUI (0.2) has very slow versions of hash code and equals methods for performance sensitive classes like AttributeValuePair. It still showed the benefit of the optimisations but made it 10 times slower to answer the queries. The modified version of these classes is only available in the JRDF subversion repository.