Assessing metric structures on GPGPU environments

Abstract

Similarity search consists on retrieving objects within a database that are similar or relevant to a particular query. It is a topic of great interest to scientific community because of its many fields of application, such as searching for words and images on the World Wide Web, pattern recognition, detection of plagiarism, multimedia databases, among others. It is modeled through metric spaces, in which objects are represented in a black-box that contains only the distance between objects; calculating the distance function is costly and search systems operate at a high query rate. Metrical structures have been developed to optimize this process; such structures work as indexes and preprocess data to decrease the distance evaluations during the search. Processing large volumes of data makes unfeasible the use of such structures without using parallel processing environments. Technologies based on multi- CPU and GPU architectures are among the most force due to its costs and performance.