Solving Multiple Queries through a Permutation Index in GPU

Authors

  • Mariela Lopresti LIDIC. Universidad Nacional de San Luis.
  • Natalia Miranda LIDIC. Universidad Nacional de San Luis.
  • Fabiana Piccoli LIDIC. Universidad Nacional de San Luis.
  • Nora Reyes LIDIC. Universidad Nacional de San Luis.

DOI:

https://doi.org/10.13053/cys-17-3-1560

Keywords:

Metric space, approximate similarity search, permutation index, high performance computing, GPU.

Abstract

Query-by-content by means of similaritysearch is a fundamental operation for applications thatdeal with multimedia data. For this kind of queryit is meaningless to look for elements exactly equalto the one given as query. Instead, we need tomeasure dissimilarity between the query object and eachdatabase object. The metric space model is a paradigmthat allows modeling all similarity search problems.Metric databases permit to store objects from a metricspace and efficiently perform similarity queries overthem, in general, by reducing the number of distanceevaluations needed. Therefore, the goal is to preprocessa particular dataset in such a way that queries can beanswered with as few distance computations as possible.Moreover, for a very large metric database it is notenough to preprocess the dataset by building an index,it is also necessary to speed up the queries via highperformance computing using GPU. In this work weshow an implementation of a pure GPU architecture tobuild a Permutation Index used for approximate similaritysearch on databases of different data nature and tosolve many queries at the same time. Besides, weevaluate the tradeoff between the answer quality andtime performance of our implementation.

Downloads

Published

2013-10-01