Marius: Learning Massive Graph Embeddings on a Single Machine
Talkat 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI '21) · 2021-07-14..07-16·
summary
A system called Marius trains billion-edge graph embeddings on a single machine by combining partition buffer, asynchronous I/O, pipelining and a new beta ordering that minimizes swaps. It runs the Twitter billion-edge run in 3.5 hours versus 35 hours for DJLk, is 3.7× faster than PyTorch-BigGraph on Freebase‑86M while matching accuracy, and scales to >1B edges and ~550 GB of parameters on a 16 GB GPU / 64 GB CPU machine. Source is open source.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:08Learning Graph Embeddings
- 2:10Key Challenge: Data Movement
- 3:40Scaling to Large Graphs: marius
- 5:38Processing Partitions
- 6:12Edge bucket orderings and I
- 7:47Buffer aware Edge Traversal Algorithm (BETA)
- 9:06Experimental evaluation
- 9:43Accuracy and Runtime Comparisons
- 10:20Compared Orderings
- 11:00Buffer-aware Edge Traversal Algorithm (BETA)
- 12:00Conclusion & Future Work
related talks
43mBuilding Billion Node Graphs for Machine Learning
50mBig Data on Smallish Machines
16mDeep Stream Dynamic Graph Analytics with Grapharis
12mMake ML outcomes roar with graph neural networks on TigerGraph and AWS (PRT097)
39mThorny path to the Large Scale Graph Processing
27mData @scale Seattle - Yucheng Low
39mThorny path to the Large Scale Graph Processing
29mGraph Neural Networks + LLM on