TopoOpt: Co-optimizing Network Topology and Parallelization for ML

← Back to the timeline

Year: 2023  ·  Venue: USENIX NSDI  ·  Authors: Weiyang Wang, Moein Khazraee, Zhizhen Zhong, et al.

Find the paper on Google Scholar

Abstract

Reconfigure the physical topology to match each ML job’s communication pattern. The network stops being a fixed substrate and becomes part of the training system.

My Notes

I haven’t written up my notes for this one yet — this page is a placeholder that I’ll fill in as I read.