r3d3: Optimized Query Compilation on GPUs

Abstract

Query compilation is an effective approach to improve the performance of repeated database queries. GPU-based approaches have significant promise, but face difficulties in managing compilation time, data transfer costs, and in addressing a reasonably comprehensive range of SQL operations. In this work we describe a hybrid AoT/JIT approach to GPU-based query compilation. We use multiple optimizations to reduce execution, compile, and data transfer times, improving performance over both other GPU-based approaches and CPU-based query compilers as well. Our design addresses a wide range of SQL queries, sufficient to demonstrate the practicality of using GPUs for query optimization.