Exploiting GPUs for fast intersection of large sets

Abstract

The main focus of this work is on large set intersection, which is a pivotal operation in information retrieval, graph analytics and database systems. We aim to experimentally detect under which conditions, using a single graphics processing unit (GPU) is beneficial over CPU techniques and which exact techniques are capable of yielding improvements. We cover and adapt techniques initially proposed for graph analytics and matrix multiplication, while we investigate new hybrids for completeness. We also explain how we can address set containment joins using the same techniques. The comprehensive evaluation highlights the main characteristics of the techniques examined when both a single pair of two large sets are processed and all pairs in a dataset are examined, while it provides strong evidence that state-of-the-art set containment stands to significantly benefit from advances in GPU-enabled set intersection. Our results reveal that there is no dominant solution but depending on the exact problem and the dataset characteristics, different techniques are the most efficient ones.