DFG project part of the priority programme 2377: https://spp2377.uos.de/

Abstract:

Code generation for GPU database systems creates fast runtime code, but suffers from expensive compilation and requires all modules to be compiled before execution. This proposal has three objectives: First, we investigate how the NVIDIA Runtime Compiler (NVRTC) accelerates compilation for code-generating GPU database systems. Using NVRTC, we demonstrate just-in-time (JIT) compilation for a tuple-at-a-time (push-based) query execution model. Second, we measure the suitability of morsel-driven parallelism for GPU database systems. This will enable the development of code-generating GPU database systems for adaptive query execution. Third, we aim for a seamless integration of GPU database systems with machine learning tools.