Using graphics processors for high performance ir query processing

Web sear h engines are fa ing formidable performan e hal-lenges as they need to pro ess thousands of queries per se ond over billions of do uments. To deal with this heavy workload, urrent engines use massively parallel ar hite tures of thou-sands of ma hines that require large hardware investments. We investigate new ways to build su h high-performan e IR systems based on Graphi al Pro essing Units (GPUs). GPUs were originally designed to a elerate omputer graphi s ap-pli ations through massive on- hip parallelism. Re ently a number of resear hers have studied how to use GPUs for other problem domains in luding databases and s ienti omput-ing [2, 3, 5℄, but we are not aware of previous attempts to use GPUs for large-s ale web sear h. Our ontribution here is to design a basi system ar hite ture for GPU-based high-performan e IR, and to des ribe how to perform highly eÆ-ient query pro essing within su h an ar hite ture. Prelimi-nary experimental results based on a prototype implementa-tion suggest that signi ant gains in query pro essing perfor-man e might be obtainable with su h an approa h. Categories and Subje t Des riptors: H.3 [INFORMA