EVPN Flooding Bug, Network-Aware Scheduling Proven, Microsoft Breaks from OpenAI
A silent flooding bug in Arista EOS EVPN centralized routing, a formal proof that ignoring network topology in AI scheduling is mathematically unjustifiable, and Microsoft shipping its first in-house reasoning model while open-sourcing MRC transport tooling.
Welcome to Amaze Networks for Wednesday, June third. Remember Monday when we said connectivity is the next AI bottleneck — and then Marvell put Jensen Huang on stage to prove it? Today we have a paper that makes the same argument mathematically, specifically for the scheduler layer, and the performance gap they measured is not a rounding error.
Twenty-one percent reduction in Time to First Token. Proven formally, not just empirically. And the authors show there's no upper bound on how bad it gets as context length grows. But that's our second story. Let's start with the one that had the highest RSS score this morning — Ivan Pepelnjak dropped another post in his EVPN centralized routing series, and this one has an Arista-specific gotcha that I am confident is causing silent flooding in production networks right now.
This series has been building toward something useful. What does this installment add?
New episodes, every weekday.
Amaze Networks drops at 4 AM CT, Monday through Friday. Spotify and Apple Podcasts submissions in progress.