Reducing I/O Cost of Similarity Queries by Processing Several at a Time

Current research on high-dimensional indexing structures for multimedia retrieval is focused on obtaining optimal performance for single k-NN queries. Within this paper, we argue for parallel processing of multiple nearest neighbor queries. This is beneficial when there is heavy load on the server and efficiency is important. We argue that query performance for parallel queries is an important parameter in the evaluation of the usefulness of indexing structures. We show that by processing multiple queries simultaneously we can dramatically reduce the cost of each single query involved. We support these findings via experiments on VA- and inverted VA-files, an improved variant of VA-files.