Large-scale Graph Learning in a Single Machine
We are developing Marius to make the use of deep learning models over billion-scale graphs easier, faster, and cheaper. Marius focuses on the key bottleneck of applying deep learning over over large-scale graph data: data movement. Marius addresses this bottleneck with a novel data flow architecture that maximizes resource utilization of the entire memory hierarchy (including disk, CPU, and GPU memory).
Marius is under active development. It is funded by DARPA under the ASKE AI Exploration (AIE) program and released under the Apache-2.0 License. You can learn more about Marius from our recent OSDI 21 and VLDB 21 talks.