Engineering Facebook's Data Center Network: Challenges and Limitations
summary
Facebook redesigns data center networking to scale past terabits of machine-to-machine traffic by shifting from vertically limited CSW boxes to a horizontally scalable fabric that can span a floor as one cluster, deployable from 4:1 to 1:1 oversubscription. Key pieces: a Wedge top-of-rack switch with onboard microserver and BMC running fboss Linux OS, and a 12-board spine chassis (8 front, 4 fabric). Production tests run in-site, optical upgrade path targets 40→100→200→400 Gb reuse of fiber, and one engineer monitors alerts.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:48Our For Post Architecture
- 5:17Control Configuration Monitoring
- 6:37Topology
- 7:55TopIraq
- 8:55Fbus
- 9:31Spine Switch
- 10:33Open Platform
- 12:00Whats Next
- 14:28Operational Complexity
related talks
37mDisaggregating the Network: Switching as a Service
11mRunning BGP in Data Centers at Scale
17mMeta’s Network Journey to Enable AI
46mLessons learned from monitoring a large production network
18mKeynote: Enterprise Data Center Migration to Hyperscale – Zaid Ali Kahn
20mFiber Networks Innovations for AI at Scale
14mRDC: Energy-Efficient Data Center Network Congestion Relief with Topological
39mLoad Balancing Building Blocks