Compiling Query Plans to Efficient GPU Programs
Master's Thesis
Henning Funke
Abstract
Parallel processing of database operators on GPUs offers dramatic speedups over CPU processing. Still the very high arithmetic throughput of the GPU cores makes data movement the limiting factor for throughput. Especially operator-at-a-time approaches suffer from a large memory load caused by intermediate materialization.
Executing query plans as pipelined operator sequences allows to reduce the memory intensity of processing. However many GPU operators contain parallel data dependencies that prohibit a fully pipelined execution. In this work, we show ways to generate kernel code for efficiently pipelined query execution on GPUs.