GPU based accelerator for RankBoost in web search engines

摘要

The general ranking problem has widespread applications including commercial search engines. RankBoost is an efficient ranking algorithm for combining preference in these areas. But it is not widely used because of its long training time. Graphics Processing Units(GPUs) have become powerful parallel processing tools for general purpose computing. In this paper, we use CUDA compatible GPU to accelerate RankBoost training procedure. Based on the parallel architecture of GPU, we propose two mapping schemes: One-Feature-One-Thread (OFOT) and One Feature-Multiple-Thread (OFMT). Different training data-sets lead to different speedups using our mapping schemes. For training data sets from a commercial search engine, the OFOT is better, achieving a 30× speedup; for random data, the OFMT is better achieving a 60× speedup.