Maybe-Holistic SDN Model?

One of my biggest frustrations about SDN has been the lack of a complete top-to-bottom architecture.  All of the focus seems to be on the SDN Controller, and that’s an element that is a little functional nubbin that lies between two largely undefined minefields—the lower-layer stuff that provides network status and behavior and the upper-layer element that translates service requests into routes based in part on that status/behavior.  Now we may have at least a step toward a vertically integrated model.

Pica8 has announced an SDN architecture for the data center that’s vertically integrated to the point that it looks a lot like a cloud-provisioning model (Quantum) in terms of the functional boxes.  There’s an open switch abstraction (OVS) linked with a network OS and a hardware layer that adapts the central logic to work with various devices, including “white box” generic switches.  The current Pica8 announcement is focusing on the application of this architecture to the problem of data center networking, not so much segmentation (though obviously you can do that) but to traffic engineering and creating efficient low-latency paths by meshing switches rather than connecting them into trees (the current practice with Ethernet) or turning them into fabrics.

This model of SDN application could be one of the sweet spots for SDN because it’s addressing a very specific issue—that cloud or even SOA data centers tend to generate more horizontal traffic without becoming fully connective in a horizontal sense.  In SOA, for example, you have a lot more intercomponent traffic because you have deployed separate components, but that traffic is still likely less than the “vertical” flows between components and users or components and storage systems.  In traditional tree-hierarchy switched networks, horizontal traffic might have to transit four or five layers of switches, which greatly increases the delay and the overall resource load.  Fabrics, which provide any-to-any non-blocking switching, waste some of that switch capacity by automatically supporting paths that have no utility, or are not even contemplated.

The Pica8 architecture is also interesting,  as at least offers the potential to combine real telemetry from the network and real service requests from data center/cloud software to create paths.  As I noted earlier, there are few models of SDN that provide the vertical stack even in limited form, so it’s heartening to see something come out.  The problem is that the model of the data center, while it may offer sweet-spot early positioning, doesn’t expose the full set of value propositions or issues.

Every data center doesn’t need a fabric or mesh.  While we might want to believe that VM growth (private cloud or virtualization) or other architectural factors would change this, the fact is that data center networking needs are set more by total application traffic than anything else, and moving around where applications are hosted doesn’t impact traffic very much.  A major increase in the application traffic would imply a much larger investment in IT resources, and it’s clear from the earnings reports that kind of growth isn’t happening.  It may, if our model of point-of-activity empowerment matures, but not yet.  Thus, data centers are not necessarily under a lot of growth pressure.

The dynamism of future applications will generate network agility requirements before it will generate traffic, but the question that Pica8 and everyone else will have to answer is how those requirements move out of the data center.  A rope staked to the ground on one end can move only in a circle.  If the edge of the network, the client devices and users, are still doing the same old thing then the changes in the data center will dissipate as they move toward the edge and the total network won’t change much.  Not much dynamism.  Even a zillion mobile clients hooking up to enterprise apps really doesn’t do anything that SSL connections to a web server for worker or even customer access doesn’t do.  You need a new application model that drives a new connection model, one that takes SDN out of the data center and rolls it all the way to the edge.

We need to be watching Pica8 at this point to see how it plans to support this sort of migration.  We also need to see how well it will address the metro-cloud opportunity that is the service provider equivalent of that enterprise network drive I called point-of-activity empowerment.  It’s a promising start but we need more progress to call it a win.

Leave a Reply