Efficient Data Management for GPU Databases

General purpose GPUs are a new and powerful hardware device with a number of applications in the realm of relational databases. We describe a database framework designed to allow both CPU and GPU execution of queries. Through use of our novel data structure design and method of using GPUmapped memory with efficient caching, we demonstrate that GPU query acceleration is possible for data sets much larger than the size of GPU memory. We also argue that the use of an opcode model of query execution combined with a simple virtual machine provides capabilities that are impossible with the parallel primitives used for most GPU database research. By implementing a single database framework that is efficient for both the CPU and GPU, we are able to make a fair comparison of performance for a filter operation and observe speedup on the GPU. This work is intended to provide a clearer picture of handling very abstract data operations efficiently on heterogeneous systems in anticipation of further application of GPU hardware in the relational database domain. Speedups of 4x and 8x over multicore CPU execution are observed for arbitrary data sizes and GPU-cacheable data sizes, respectively.