high performance computing on graphics processing units: hgpu.org

Modern graphics cards bundle high-bandwidth memory with a massively parallel processor, making them an interesting platform for running data-intensive operations. Consequently, several authors have discussed accelerating database operators using graphics cards, often demonstrating promising speed-ups. However, due to limitations stemming from limited device memory and expensive data transfer, GPUaccelerated databases remain a niche technology. We suggest a novel approach: Using the graphics card as a co-processor during query optimization. Query optimization is a compute-heavy operation that requires only minimal data transfer, making it a well-suited target for GPU offloading. Since existing optimizers are typically very efficient, we do not suggest to simply accelerate them. Instead, we propose to use the additional resources to leverage more computationally involved optimization methods. This approach indirectly accelerates a database by generating better plan quality. As a first step towards GPU-assisted query optimization, we present a proof-of-concept that uses the graphics card as a statistical co-processor during selectivity estimation. We integrated this GPU-accelerated estimator into the optimizer of PostgreSQL. Based on this proof-of-concept, we demonstrate that a GPU can be efficiently used to improve the quality of selectivity estimates in a relational database system.