Fabric design in data centers

There are three main reasons for deploying the control plane on the hosts attached to the fabric in DCs:

• Faster Down Detection. Let the hosts themselves advertize correct information and the routes will be automatically be withdrawn on failure.
• Load Sharing. Default route on dual-homed hosts makes the hosts blind to the network conditions, which leads to equal load-sharing only. Equal load sharing, however, may not be ideal in all situations. Running routing makes injecting more intelligence in the load sharing possible. Like one link for data and one for voice, but keep the failover possible.
• Traffic Engineering. Some traffic may prefer one path over another in the fabric. For this traffic, the only way is to run the control plane on the hosts to inject a lable/outer header onto the packet.

Normally, we need to separate customer routes from infrastructure routes in SP networks to have separated failure and security domains, which also dividing telemetry data that allows for faster troubleshooting.

The same principle should also be applied to a DC fabric by using two different routing protocols for the underlay and overlay as the simplest way. For example, ISIS for the underlay and eVPN+BGP for the overlay.