Optimize First, Buy Later: Analyzing Metrics to Ramp-up Very Large Knowledge Bases

As knowledge bases move into the landscape of larger ontologies and have terabytes of related data, we must work on optimizing the performance of our tools. We are easily tempted to buy bigger machines or to fill rooms with armies of little ones to address the scalability problem. Yet, careful analysis and evaluation of the characteristics of our data-using metrics-often leads to dramatic improvements in performance. Firstly, are current scalable systems scalable enough? We found that for large or deep ontologies (some as large as 500,000 classes) it is hard to say because benchmarks obscure the load-time costs for materialization. Therefore, to expose those costs, we have synthesized a set of more representative ontologies. Secondly, in designing for scalability, how do we manage knowledge over time? By optimizing for data distribution and ontology evolution, we have reduced the population time, including materialization, for the NCBO Resource Index, a knowledge base of 16.4 billion annotations linking 2.4 million terms from 200 ontologies to 3.5 million data elements, from one week to less than one hour for one of the large datasets on the same machine.