AIN: A Blueprint for an All-IP Data Center Network
summary
Proposes replacing virtual switches with a virtual IPv6 router in each hypervisor, mapping VM addresses into a 128-bit network space via tenant/network tags plus a consistent hash that picks a home hypervisor for each VM. Uses OSPF areas and ECMP for scalable routing (example sizing shows modest LSA rates for 10k–100k pods), mobility via proxy Mobile IPv6 with GRE tunnels to current hosts, and broadcast-only emulation by hypervisor encapsulation; claims full compatibility with existing apps and current protocols/hardware.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:29Cloud Network Requirements
- 1:19Which Layer to Focus on?
- 4:49Two Address Spaces
- 5:43Example: Host to Network Address Space
- 7:01Home Hypervisor
- 8:40Typical Data Center Network Architecture
- 10:29Mapping to OSPF
- 11:10Consequences of OSPF
- 12:16Broadcast Support
- 13:03Roadmap
- 15:28Distributed PMIP for Mobility in AIN
- 16:02Consequences of PMIP
- 17:03Multitenancy
- 17:57Related Work
- 19:11Summary
related talks
22mLayer 3 networking with BGP in hyperscale DCs
29mLiving on the Edge with DCN Networking
38mHow can a network be "virtual"?
26mIPv6 First, Not Just Ready: Kubernetes Without IPv4 Using Cilium
22mTransparent MultiNIC Routing for Large AI Models
27mLeveraging Cilium and SRv6 for Telco Networking
27mNetwork Multi-tenancy in Xen-based Clouds
22mConnecting virtual machines in the cloud age