Abstract

Advances in GPU technology have frequently driven innovations in database query processing. The recent incorporation of native support for ray tracing (RT) in modern GPUs has, as expected, spurred interest in its application to relational database workloads — in particular, RT has been advocated for column indexing. We examine this premise here and conversely argue that it is the classical rasterization pipeline, rather than RT, which is to be preferred for such applications. The rationale is that rasterization uses arithmetic value comparisons, instead of the computationally heavy ray-triangle geometric intersections of RT, for data search. To substantiate this claim, we design RasterScan, a purely rasterization-based indexing approach, that adheres to the data model employed by RT-based methods. Our evaluation over a variety of large datasets demonstrates that RasterScan consistently and substantively outperforms its RT counterparts, often achieving order-of-magnitude speedups with regard to both index build and search times.