Hi,
It’s very likely that it will be technically possible to connect end systems to the spine layer, but whether it’s the “right” thing to do will depend upon what you’re trying to build and the reasons you’re building it.
We know that spine-leaf networks are simply a term used to describe a Clos network, and that the concept of a Clos network came from the telephony world as a way of overcoming the high cost of building large switching fabrics. From Wikipedia:
“Clos networks are required when the physical circuit switching needs to exceed the capacity of the largest feasible single crossbar switch.”
Or as Russ White / Denise Donohue put it in The Art of Network Architecture: Business-Driven Design “The general idea behind a Clos fabric is that a network capable of carrying high traffic rates can be built from components that individually support only lower packet switching speeds.”
What this means is that instead of building a data centre network comprised of two large chassis based switches at the aggregation as we’ve done in the past, we can build it using low cost, low density switches at the spine, with similar low cost, low density switches at the leaf. The network is flatter and wider than traditional aggregation-access networks, and using Layer-3 Equal Cost Multi Path, we get to load balance traffic across all the possible paths.
There’s a research paper, A Scalable, Commodity Data Center Network Architecture from 2008 written by amongst others, Amin Vahdat, who’s now the Technical Lead for Networking at Google, that goes into this approach in detail. In this paper the spine is considered as the middle stage of the Clos fabric, and so only used to connect leaf switches.
I quite like the way that Big Switch Networks consider a spine-leaf fabric. If you look at their Big Cloud Fabric™ they equate a spine-leaf network to a "big switch" with the spine layer analogous to the switch backplane:
“The BCF (P Fabric) architecture maps readily to a traditional redundant chassis switch architecture:
- Supervisor Card(s) <--> Controller Cluster
- Chassis Backplane <--> Spine Switches
- Line Card(s) <--> Leaf Switches”
And show the following diagram:
So if you consider a spine-leaf fabric in these terms then clearly you wouldn’t connect end systems to the spine layer, in the same way you don’t connect them to the backplane of a chassis based switch.
This is perhaps a somewhat purest view of a spine-leaf network, and really considering spine-leaf networks for large scale deployments, but spine-leaf is a little like Software Defined Networking (SDN) in that it can mean different things to different people.
So I guess the very short answer is, that it depends J
Regards
