Current Internet Exchange (or any other business exchanges that need to
provide ad-hoc peering services) architecture is mostly composed to big
chassis routers/switches, those big boxes are configured to provide a big L2
-fabric, either pure L2 or different flavors of MPLS L2 VPNs, problem with
this design is, per-port cost of big switches are high, network can get
overly complicated in control plane, big boxes tend to less stable than
fixed configuration because the software is much more complicated, and this
solution does not scale linearly.
People have been using spine-leaf topology with fixed 1RU/2RU small switches
in data centers for a while, Facebook extended this topology to a new level
in their hype-scale Iowa data center -- without using a single big box,
this topology provides multiple layers of fault tolerance, besides, the BGP
controller can do lot of interesting stuff.
Now can we not use the same idea in Exchange environment? we have a solid L3
fabric, we then provide VXLAN connectivity on leaf switches, this should
fix all shortcomings of current Exchange architecture.
Thoughts? maybe Exchange environments will never need that level of
scalability?