Title:	WALP Project: ORB DB: Neural Algorithmic Reasoning for Graph Databases
DNr:	Berzelius-2024-362
Project Type:	LiU Berzelius
Principal Investigator:	Paris Carbone <parisc@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2024-10-01 – 2024-12-01
Classification:	10201
Homepage:	https://orbdb.github.io/
Keywords:

Abstract

In this Wallenberg Launchpad project we validate the hypothesis that learned computation can be leveraged within graph processing query plans to enhance output and speed. Our biggest challenge currently is to train graph neural networks on billion-vertex graphs. Primarily the Query2Box model can take more than 10 days currently to train with commodity GPUs on large graphs such as OGB-LSC. We would like to use Berzelius to reduce this to 1 day per model. Furthermore, we would also like to request if possible bare metal access. This training process requires the most advanced GPU hardware we can find in Sweden.