Express Lane to Efficiency and Reliability: Multi-Dimensional Control in Meta’s Express..
summary
Meta’s Express Backbone is eight parallel MPLS/segment‑routing planes and grows with AI traffic; the team adds eight planes aware load balancing via Unequal Cost Multi-Path (UCMP) that uses an iterative water‑filling algorithm to compute per‑site‑pair vectors of eight weights, cutting the simulated bandwidth deficit by ~75%, and a controller‑computed Fast Re‑Route (FRR) that drops worst‑case convergence from 90s to 3s (hardware FRR sub‑second); a real migration draining 25% capacity shows ~95% reduction in bronze deficit.
AI-written summary · about
chapters (9)
- 0:00Express Backbone at Meta
- 1:22EBB architecture challenges
- 4:09Cross-plane capacity problem
- 6:12UCNP approach solution
- 7:55Path protection limits
- 9:40EBB Fast Route restoration
- 12:08Production results gains
- 13:10Migration case study
- 14:34Conclusion future-proofing
related talks
22mTransparent MultiNIC Routing for Large AI Models
13mLive in the Express Lane
28mSafe sharing of the network with eBPF
17mTraffic Engineering for AI Training Networks
13mAUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
18mLive Group Q&A Session
17mUnlocking ECMP Programmability for Precise Traffic Control
23mMinimal Rewiring: Efficient Live Expansion for Clos Data Center Networks