#IoT #OpenStack #SDN
When people write about software-defined architectures being "disruptive" to the network they're doing a bit of a disservice to just how much change is occurring under the hood in the engine that drives today's businesses. The notion of separating control and data planes is superficial in that it describes a general concept and it isn't really all that radical a change, if you think about it.
The control and data planes have always been separate. We have, since the need for web-scale networks came about, implemented separate topological (and usually physical) networks specifically for the purpose of segregating control traffic from the data path. The reasons for this are many: to keep management (control) traffic from interfering with the delivery of applications (and vice versa), to enable a model in which control over the critical path for applications could be secured and to ensure access to necessary control functions in the face of failure or attack.
What's new with software-defined architectures is not just the logical separation but the physical decoupling and change in component responsibility. In traditional networks there is a logically separate control plane, but it is distributed; it resides on each physical component. In software-defined networks it is physically separate but it is centralized; control responsibility resides in a single component, the "controller".
Now, OpenStack and emerging models for scaling systems that will be responsible for managing communication with and for the Internet of Things (like MQTT) use a similar control model, but it's not as active as a software-defined architectural model, it's more a passive model. That's the nature of PubSub imposing itself on the network.
PubSub (publish / subscribe) is a familiar model to application developers. It's a middleware staple that's been used for a very long time to distribute messages to a variable set of systems. In a nutshell, PubSub is based on the notion of there existing a "queue" to which authorized components can publish events or messages of interest and to which interested components an subscribe. Events or messages have a life (like a TTL) and eventually expire. In the interim, it's expected that subscribed components check the queue for messages periodically. They poll for events or messages.
The queue itself is much like a switch or router's queue, except messages in the queue are duplicated until the TTL runs out and the message expires.
PubSub is passive; that is, it does not actively distribute messages. It merely serves as a kind of centralized repository, making available to those components that need it access to relevant information about the state of applications and/or the network.
OpenFlow-based SDN, by contrast, is active. That is, it not only serves as a centralized repository for the state of applications and/or the network, but it actively distributes messages to components based on events. For example, a controller might receive a message from another system indicating the launch of a new application instance. That event triggers a series of actions on the controller that includes informing the affected network components of configuration changes. In a passive, PubSub model, the network components themselves might be polling for such an event and, upon receiving one, would initiate the appropriate configuration changes themselves.
We can simplify the description of the differences even more: a controller-based architecture uses a push model, while a pubsub-based architecture uses a pull model. What this doesn't illustrate well is that in a push model, the centralized controller must know how to communicate the desired changes to each and every component it is controlling. That's one of the reasons original models standardized on OpenFlow and were tightly focused on L2-4 stateless networking. It could be easily standardized down to a common forwarding table.While different components might internalize that differently, the basic information was always the same: IP addresses, ports and actions.
As we move up the stack into L4-7 stateful networking, however, this model becomes more burdensome because of the complexity of rule sets and differences in policy models across such a broad set of networking domains. Hence the plug-in support in controllers like OpenDaylight for "other" control protocols. But the basic premise of the model remains the same, regardless of the control protocol: the centralized controller dictates the changes to all components. It pushes those changes to the network. Both control and execution are centralized. The controller tells components to change their configuration.
PubSub centralizes control but decentralizes execution. The control plane is still centralized; there is one authoritative system responsible for disseminating change across the network, but each individual component (or domain controller but we'll get to that in a minute) is responsible for executing the appropriate changes based on their configured policies and services. A pubsub controller never tells a component "change this now"; that's up to the individual components (or domain controller).
To make things even more confusing (and disruptive), these models may be used simultaneously. A software-defined architecture might be based on a centralized control model with domain controllers for specific networking functions (like security and application delivery) integrated via a pubsub-based model.
This is where we start seeing models that combine emerging technologies like OpenStack and SDN architectures together. OpenStack manages at the data center level, and at its heart is pubsub model that can be used by domain controllers (stateless L2-4 SDN, stateful L4-7 SDN, etc...) to receive notification of changes in the network and subsequently push those changes using the appropriate control protocols to the components it is managing.
Needless to say, the term "disruptive" is really inadequate in describing the level of change in the network required to support either models (or both). Both require significant changes not just to the network itself but the way in which the network is fundamentally provisioned and managed. It's not just a new CLI or management console, these models dramatically change the design and management of networks.