Collision Counting Locality-sensitive hashing on commodity GPU cluster

Abstract

Porting program to run on GPU has been a trend for years. However, GPU device memory is always the obstacle of making importation of data intensive program due to its small size and relatively expensive memory. LSH is an approximate nearest neighbor search method and known to be data intensive. C2LSH, as a scheme of LSH, also requires hundreds of seconds of time to complete a batch query. This research is aiming to develop the first C2LSH GPU implementation as well as its commodity GPU cluster version. With limited GPU device memory on each node, we managed to process a dataset that could never fit in the memory of a single GPU card, by our implementation of MPI-CUDA mixed programming. Furthermore, by running experimental studies to prove that we may match the computational power of a high-end card with commodity GPU cluster with a lower cost.