Matryoshka: Realizing Hyperscale Data Center Network Design for the AI Era
summary
Matryoshka compiles high-level Thrift network intents into vendor-specific switch configs, modeling topology, hierarchical IP allocation, BGP and routing policies end-to-end and translating to per-vendor configs via converters. It operates statelessly, recompiling from scratch for greenfield and brownfield changes, and drives nearly 900 DCNs across 18 designs including a 100K-GPU supercluster. Outcomes: 95.4% first-attempt config pass and 96.4% daily build success, database transaction time cut by 70–99% via incremental-diff updates, and a generic switch-config abstraction that hides vendor heterogeneity.
AI-written summary · about
chapters (12)
- 0:00Matryoshka introduction
- 1:42Meta's network evolution
- 3:06Hyperscale config challenges
- 4:55What is Matryoshka
- 6:40Matryoshka's 3-layer architecture
- 9:08Matryoshka ecosystem
- 9:49100K GPU stress test
- 10:53Performance results
- 12:18Scalability and stability
- 13:016 years lessons learned
- 14:47Evolving challenges
- 15:07Key takeaways
related talks
17mMeta’s Network Journey to Enable AI
10m10x Network Design Engineer
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
18mKeynote: Enterprise Data Center Migration to Hyperscale – Zaid Ali Kahn
10m10x Network Design Engineer
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
20mAnatomy of @Scale Machine Learning Network Operations
31mLive Panel: Scale-Out Networking for AI Supercomputers