An empirical evaluation of a distributed clustering-based index for metric space databases

Similarity search has been proved suitable for searching in very large collections of unstructured data objects. We are interested in efficient parallel query processing under situations of continuous streams of queries as in search engines. A number of sequential index data structures for this purpose have been proposed so far. This paper focuses on one representative of a class of these data structures, namely one based on clustering for which we evaluate different ways of distributing the index to support parallelism on a set of processors. Our study reveals that the intuitive method for both data distribution and model of computing are not efficient in practice. The best results are obtained with a strategy that appears to be more costly in construction but we show that in practice this cost is not significant.