This paper presents SPOWL, which uses Spark to perform OWL reasoning over large ontologies. SPOWL acts as a compiler, which maps axioms in the T-Box of an ontology to Spark programmes, which will be executed iteratively to compute and materialise a closure of reasoning results entailed by the ontology. Such a closure is then available to queries which retrieve information from the ontology. Compared to MapReduce, adopting Spark enables SPOWL to cache data in the distributed memory, to reduce the amount of I/O used, and to also parallelise jobs in a more flexible manner. We further analyse the dependencies among the Spark programmes, and propose an optimised order following the T-Box hierarchy, which makes the materialising process terminate with minimum iterations. Moreover, SPOWL uses a tableaux reasoner to classify the T-Box, and the classified axioms are complied into Spark programmes which are directly related to the ontological data under reasoning. This not only makes the reasoning by SPOWL more complete, but also avoids processing unnecessary rules, as compared to evaluating certain rulesets adopted by most state-of-the-art reasoners. Finally, since SPOWL materialises the reasoning closure for large ontologies, it processes queries retrieving ontology information faster than computing the query answers in real time.