Using semantic technologies the materialization of implicit given facts that can be derived from a dataset is an important task performed by a reasoner. With respect to the answering time for queries and the growing amount of available data, scaleable solutions that are able to process large datasets are needed. In previous work we described a rule-based reasoner implementation that uses massively parallel hardware to derive new facts based on a given set of rules. This implementation was limited by the size of processable input data as well as on the number of used parallel hardware devices. In this paper we introduce further concepts for a workload partitioning and distribution to overcome this limitations. Based on the introduced concepts, additional levels of parallelization can be proposed that benefit from the use of multiple parallel devices. Furthermore, we introduce a concept to reduce the amount of invalid triple derivations like duplicates. We evaluate our concepts by applying different rulesets to the real-world DBPedia dataset as well as to the synthetic Lehigh University benchmark ontology (LUBM) with up to 1.1 billion triples. The evaluation shows that our implementation scales in a linear way and outperforms current state of the art reasoner with respect to the throughput achieved on a single computing node