Software Defined Networking (SDN) has created two opposing views of what the data center network should look like. One side argues for intelligence, meaning that the core of the data center network itself should be intelligent enough to decide what data flows look like from one compute node to another on the Ethernet fabric. The other side argues for abundance, making the case that if the core has sufficient bandwidth, the intelligence can be relegated to the compute node, where powerful servers can make their own data flow decisions. Which side has the right vision?
An old adage states, “Where one stands on issues depends on where he sits.” Namely, if you sit in a building owned by a dominant networking equipment vendor that relies on very healthy sales margins, your stance on this issue is to argue for intelligence at the data center network core. Otherwise, you will argue for abundance. Let’s understand why.
The SDN argument for abundance assumes the following:
- Ethernet switching equipment is generally commoditized. Because of established standards such as MP-BGP or ISIS, equipment from various vendors can work well with each other to provide a capable data center core network.
- Due to this commoditization, it is to the data center owner’s interest to purchase the most cost effective equipment possible, granted certain minimal standards.
- If data center Ethernet resources are strained, the appropriate response is to add more switches and ports to increase capacity at a low-cost.
- Because servers are sufficiently powerful and continue along Moore’s Law’s trajectory, it makes sense to push the intelligence to the edge, where the compute nodes are.
In essence, the abundance view argues that intelligence at the data center network core is of little value and artificially boots costs. Clearly, those who argue for intelligence at the network core disagree. Their argument can be summed up as:
- Networking works best when the right switching and routing decisions are made at every level of the stack. As such, the hypervisor should be network aware, and the data center network should be aware of the hypervisor at the edge node.
- With this integrated approach, networking decisions are optimal and paths are created based on any edge node’s communication needs (speed, latency, security, etc.)
- This approach could avoid the unnecessary expense related to the extra pipes called for by the abundance argument, but no direct evidence is provided.
The implicit statement in the intelligence argument is that the entire data center networking infrastructure is to be from the same vendor and of the same product family. In other words, having an “intelligent” data center network behooves the customer to be beholden to a single vendor and extend his lock-in.
Once vendors have lock-in, they have little incentive to offer price breaks. By comparison, let’s look at computing costs trends, as reported by Mary Meeker in her annual Internet trends presentation:
So, which side has the right vision? Based on motivations and the cost implications, my assessment is that the “intelligent” data center network core argument is not all that intelligent after all. When it comes to deploying SDN, go for abundance.
Reach me on Twitter: @amir_sharif
Thanks for reading!