How we designed an OpenStack Public Cloud at Scale with Fabric Networking
summary
Nipa Cloud migrates its new Victoria cluster into production using Tungsten Fabric to replace brittle Neutron overlays after hitting scale pain at ~10–20 racks and aiming past 100 racks; architecture uses CLOS spine/leaf with BGP EVPN/VXLAN and Anycast (distributed) gateway for ECMP and east–west scale, Nova cells to localize RabbitMQ and scale compute, Ceph multi‑AZ, Glance on S3 with symmetric sync and Cinder image‑volume cache — the system is live and reduces broadcast storms and HA failures.
AI-written summary · about
chapters (20)
- 0:00Introduction and agenda
- 1:02About Nepal Cloud and its services
- 3:15Old public cloud challenges at scale
- 5:43Network layout and scaling issues
- 8:15Centralized routing bottlenecks
- 10:00Transition to spine-leaf topology
- 12:27Spine-leaf architecture benefits
- 14:43Tungsten Fabric overview and pros
- 16:10Neutron overlay network limitations
- 18:59Why switch to Tungsten Fabric
- 20:04Production deployment and hardware choice
- 21:31Tungsten Fabric integration with OpenStack
- 23:45Tungsten Fabric architecture details
- 25:11Multi-zone OpenStack and Tungsten Fabric
- 27:48Nova cell and RabbitMQ optimization
- 29:42Cinder volume and storage management
- 31:53Cross-site data replication strategy
- 33:20Tungsten Fabric networking with Neutron
- 35:07Virtual network configuration demo
- 36:53Q&A session
related talks
43mSupporting NASA Science with OpenStack
15mThe CERN Cloud Infrastructure: 5 years in 15 minutes
20mMulti-Cloud Without the Myth: What Actually Works in Practice
18mCStar Gateway: Augmenting Public Cloud Infrastructure for Heterogeneous Network...
17mMeta’s Network Journey to Enable AI
34mFishfingers and Custard: An OpenStack and K8s Story
18mKeynote: Enterprise Data Center Migration to Hyperscale – Zaid Ali Kahn
38mFacebook's Data Center Fabric