Title: DBs + GPUs: Where are we and what’s next?

Abstract: The massive parallelism and multi-TB/sec memory bandwidths offered by modern GPUs are hugely beneficial for accelerating SQL analytics queries. In this talk, I will start with a brief overview of the ongoing efforts in both research and industry to accelerate databases with GPUs. The main part of this talk will be about our two recent works on studying the performance of GPU-accelerated databases in a multi-GPU cluster set-up and single-GPU-CPU setup. I will mainly focus on the former, explaining how we build a distributed GPU-accelerated database with ML software stacks and gain massive performance improvement. Together with the latter work (MaxBench), we try to show the current performance landscape of GPU-based database systems and shed light on the key bottlenecks and promising directions for the next generation of GPU-accelerated data systems.

Slides