A virtual machine model for accelerating relational database joins using a general purpose GPU
Kevin Angstadt, Ed Harcourt · IEEE International Conference on High Performance Computing, Data, and Analytics · 2015
We demonstrate a speedup for database joins using a general purpose graphics processing unit (GPGPU). The technique is novel in that it operates on an SQL virtual machine model developed using CUDA. The implementation compiles an SQL statement to instructions of the virtual machine that are then executed in parallel on the GPU. We use the three-dimensional structure of the CUDA grid and thread model to perform a join on up to three relations at a time. Query execution results in speedups of 2 to 60 times on consumer-level GPUs depending on the size of the result set.