In a discussion I had yesterday, I noticed that the networking world still has many arbitrary boundaries. It is what we do: create boundaries where none really exist. We do this to cut a problem down to size. Yet when that itself becomes the problem, we end up with design decisions based on our boundaries. We need networking, specifically software-defined networking, to ignore most boundaries. We need to move away from terminology that imposes those boundaries upon our designs. Virtualization is about breaking silos, not imposing them. Network virtualization needs to do the same.
So, what is the problem? The problem is how we look at leaf-spine architecture, the generalized network architecture for any modern data center. This itself imposes several arbitrary boundaries. Those boundaries are there to help us understand the generalized problem, not to become the boundaries within which we work. I asked a simple question in response to a statement. That statement was that all leafs end at the rack. This is the accepted physical description of a leaf. However, it imposes an arbitrary boundary on my network, the network of the rack with its top-of-rack switching (ToR). My concern was around redundancy and availability. I would never design a system in which all my cluster nodes were in just one rack, but that requires the leaf domain to span multiple racks. I was told that in order to do this, “you need a very good reason,” implying that such a reason does not exist.
I span clusters across racks, so if one rack has an issue, I am safe with the other nodes. My east-west traffic still flows. For that, I need to trunk VLANs between two ToR devices, in essence spanning leafs between racks. In my case, the rack is not my arbitrary boundary. My cluster becomes the arbitrary boundary. Yet, the really good reason requirement is not a reason, and it is not a very good requirement. It itself becomes the silo in which we start to live. We are not breaking silos, but building them. I rather hear, “Okay, but these are the methods we have available to do that, and here are the drawbacks and benefits.” Whether I use physical or virtual networking components does not matter; what matters is the mindset.
The above are three mindsets associated with today’s networking. As you move up the stack to full network virtualization, you end up with the far right. However, most networking folks are at the far left. There is a middle ground, where we think of the cluster of systems, not the rack, as being the arbitrary boundary, not defined by the ToR. In fact, as you move to the virtual world, the concept of ToR disappears, and we end up with network virtualization.
However, while we have abstracted away ToRs, they are here to stay. They are not a boundary themselves. They are a method for reducing cabling overhead. A rack is easier to cable if the network cables to the ToR and the ToR leaves the rack. ToR can leave the rack in either east-west (other racks) or north-south (to the spine and back) topologies. When we start talking about software defined, the ToR and all associated cabling becomes nothing more than transport. We start putting leaf switches in other places mentally, but we can also put them in other places physically. We end up with a more synergistic view of networking, one that will, in the end, mesh well with network virtualization, software-defined networking, and eventually network functions virtualization.
Those of us who live in the virtual world need to think of how that maps to the physical world. Those in the physical world need to map their thoughts to the virtual world. Remember, we all have points of view. Arbitrary boundaries should be guides, not the be-all and end-all. We should not silo on those boundaries. There are reasons for everything. We need to understand and not say “you better have a good reason.” The reason does not matter: the result does. The fact that leaf-spine was designed to sell ports should not dictate how we define our leaf-spine network. ToR may be the traditional way we look at a leaf, but it does not need to be. Let us break the arbitrary boundaries and therefore break down our silos.