UC San Diego

This project focuses primarily on improving the implementation of the Hierarchical NavigableSmall World Algorithm by maintaining the hierarchical structures and certain path of query utilizing the GPU for better performance. Specifically, multiple queries could be now used and the overall time is reduced to around 83%. Such goal is achieved by investigation of parallel computation within layer nodes, batch query, and using half-precision representation which could increase query speed without affecting the actual precision and recall of the algorithm. By incorporating these modifications, we are able to improve the algorithm in for a query by a factor of 2% - 5%.