Please note: This seminar will take place in DC 1304 and online.
Justus Henneberg, PhD student
Johannes Gutenberg University
It is a promising time for GPU databases: Modern GPUs are equipped with over 100 independent SIMD units and up to 100 GB of ultra-fast memory, and can read directly from PCIe-enabled SSDs. Unfortunately, most database-related CPU algorithms are not designed to run on a GPU efficiently. We therefore need to rethink joins, indexing, storage, and query processing to fully take advantage of the GPU's capabilities, which now even include specialized hardware for matrix multiplication and ray tracing. In this talk, we will see how database indexing can be reduced to a ray tracing problem in three different ways to take advantage of this specialized hardware, and why you can still outperform a hardware-accelerated index using a GPU-optimized (software) CSS tree.
About the speaker:
Justus Henneberg is a PhD student at Johannes Gutenberg University in Mainz, Germany. He is part of the Information Systems research group led by Felix Schuhknecht. Justus' main research interest is mapping problems to specialized hardware, which led to him proposing the idea of raytracing indexing at VLDB 2024, and co-authoring a follow-up ICDE 2025 publication with Rosina Kharal and Trevor Brown from UWaterloo.
To attend this seminar in person, please go to DC 1304. You can also attend virtually on Zoom.