Online document clustering using GPUs

Abstract

An algorithm for performing online clustering on the GPU is proposed which makes heavy use of the atomic operations available on the GPU. The algorithm can cluster multiple documents in parallel in way that can saturate all the parallel threads on the GPU. The algorithm takes advantage of atomic operations available on the GPU in order to cluster multiple documents at the same time. The algorithm results in up to 3X speedup using a real time news document data set as well as on randomly generated data compared to a baseline algorithm on the GPU that clusters only one document at a time.