In a previous blog entry, I focused on the right level of abstraction to make SDN successful and adopted in the enterprise. As promised, more focus now on another question that comes up in the case of delegation of control: resource brokering.

If every application can manipulate the network configuration and request resources, how and where do you resolve conflicts? (And who resolves them?) Don’t we all always want the best service, the fastest link, the lowest latency?

Didn't we have that problem in the past with ATM and technologies like RSVP as well? So, we are back to call admission control (CAC). It won't work without something similar to that, if you are really interested in service guarantees. To traffic-engineer around it will not always work. So, IT is back in charge to somehow provide that level of mitigation and resource management so the lines of business stay within reasonable parameters as they start to request resources.

To create a private cloud, enterprises must implement a better demand management. I keep getting feedback from customers that the perception of their business lines is such that virtualized resources come for free. Even if they are trying to establish approval workflows, they don’t have charge-back capability, so every request from their business units gets fulfilled.

That results in an ongoing increase of virtual-machine workloads in the data center. An increase of 20 percent or more per year is not the exception here (not counting the additional number of workloads created while the physical-to-virtual migration is ongoing). This level of VM sprawl almost defeats the initial purpose of virtualization: server consolidation leading to savings in footprint, cooling, and power.

On top of operational complexities associated with a fully virtualized environment, a customer of ours phrased an interesting statement: “Virtualization has resulted in a generation of application developers who assume that compute resources are infinite. So, no code optimization takes place anymore”. I quote without the name to protect the innocent… but that seems to be true: Demanding 24 cores for a single VM does not seem unrealistic, even though it could be done much more efficiently.

If you combine the increasing demand from application developers with the “no-cost” attitude from the business units, you end up with a exponentially growing number of virtual workloads that have serious physical compute and storage requirements.

Something similar could happen on the network side when resource allocations are granted in an uncontrolled fashion. But we work in an interesting field where vendors are able to differentiate themselves from each other and provide tailored solutions to their customers. Watch this space.