When do you want your network to look like a single switch?
Finally. Someone has clarified why and where does it matter to have a data center fabric that looks like a single switch, and especially to whom should it look like a single switch.
That someone is François Tallet, and his last post is very interesting. And very easy to understand, too. It revolves around two basic principles:
Data center fabrics should look like a single switch from the user perspective.
Data center fabrics should look like a real network from the network admin perspective.
Here the user is the server admin, and by extension the server itself. To them, the network would ideally look like a single fabric with all around feature consistency, any-to-any connectivity and predictable latency. From the network admin perspective, though, you need troubleshooting tools at the port level, policy enforcement mechanisms, bandwidth provisioning... and a nice distribution of the control plane never hurt anybody with regards to high availability.
What about non-blocking connectivity and a consistent low-latency design? This problem has been architecturally solved for decades with Clos architectures, and data center fabrics based in new standars like TRILL (or pre-standard Cisco's FabricPath) are easily accommodated to this architecture.
That's why data center fabrics should not be a single switch, but behave like one where it matters, and be networks everywhere else. That's why I believe "Borg" architectures like the "Stratus-pretend" that's QFabric have only half of the story covered.
