GPU-Accelerated Block-Max Query Processing

In this paper, we propose a method for parallel top-k query processing on GPU(s). We employ a novel partitioning strategy which splits the posting lists according to document ID numbers. Individual GPU threads simultaneously perform top-k query processing within their allocated subsets of posting lists, the results of the query are merged to give the final top-k results. We further design a CPU-GPU cooperative query processing method, where a majority of queries involving shorter posting lists are processed on the GPU side. We experiment with AND, OR, WAND, and Block-Max WAND (BMW) queries, with experimental results showing a promising improvement in query throughput, particularly in the case of BMW queries.