Scalable Top-K Query Processing Using Graphics Processing Unit

Abstract

Top-K query processing is one of the fundamental and the most performance-deciding components in Web search engines. A number of techniques such as dynamic pruning have been proposed to reduce the query processing time on CPU. However, it has become increasingly difficult to further improve Top-K query processing’s efficiency without hurting its effectiveness. On the other hand, Graphic Processing Unit (GPU), a powerful computing accelerator on almost every computer today, is barely tapped in Web search engines. The biggest challenge to accelerate top-K query processing on GPU is that the parallel nature of execution model of GPU prevents many CPU top-K query processing optimizations from being directly ported to GPU. GPU with hundreds of cores is ideal for applications with massive parallelism, which is not readily available in existing CPU-oriented top-K query implementations.

This paper exploits the GPU computation power for top-K query processing. In particular, we propose a new domain-specific parallelization framework to utilize GPU to parallelize it. The proposed framework is general enough for both disjunctive and conjunctive query processing modes. Experiments on TREC collections show that our proposed GPU top-K query processing framework is able to improve the query processing time by a factor of 7 when compared with state-of-the-art dynamic pruning methods for the disjunctive mode and by a factor of 6 when compared with the conjunctive mode. Our results show that our GPU top-K query processing framework is faster than previously known GPU baseline method. In particular, our framework is shown to be more scalable and efficient than the CPU and GPU baselines when K is large.