This paper is aimed at web ontology language (OWL) authors and tool developers. These people know that SROIQ(D), the logic that underpins OWL, has a high worst-case complexity (N2Exptime), and that this shows up in some practical examples. Hence, divide and conquer (splitting the logic into modules) seems like an obvious step. This modularity might help in numerous ways, notably in both reducing the number of subsumption tests that need to be performed and in reducing their difficulty (as there are fewer axioms). In practice this is not always the case, and the authors had earlier observed a case where classifying a module took longer than classifying the whole ontology.

For their experiments, the authors used four reasoners: HermiT 1.3.8, Pellet 2.3.1, JFact 1.2.3, and FaCT++ 1.6.3, and explain why they couldn’t use Konclude. They had 330 OWL full ontologies to test with, 240 of which were processed by all four reasoners. Of these, 152 did not trigger any subsumption tests. The comparatively small number of medium (100 to 1000 ms) or hard (10 to 100 s) subsumption tests, nearly all negative, dominated the contribution of subsumption to the overall running time. But that contribution was over 40 percent of the running time for less than 10 percent of the ontologies. There are many other more detailed conclusions that will interest reasoner developers, but the overall lesson is what it always has been in performance engineering: measure first and don’t trust your gut reaction.