GPU-based parallelization of QuickScorer to speed-up document ranking with tree ensembles

Scoring documents with learning-to-rank (LtR) models based on large ensembles of regression trees currently represents one of the most effective solutions to rank query results returned by large scale Information Retrieval systems. However, such scoring models are very complex, and when deployed in real Web Search Engine infrastructures they are constrained within strict time budgets. This calls for very fast and efficient solutions, able to exploit all the computational resources offered by a given system. This paper investigates the opportunities offered by modern graphic cards (GPUs) to efficiently exploit LtR complex models based on trees ensembles to rank documents. To this end we propose GPUScorer, a GPU-based parallelization of the state-of-the-art algorithm QuickScorer to score documents with tree ensembles. GPUScorer takes advantage of the huge computational power of GPUs to perform tree ensemble traversal by evaluating multiple documents simultaneously. We provide a concise experimental evaluation, and show that GPUScorer is able to achieve speedups up to 32x over the sequential version of QuickScorer