IoT: Beyond SDN

Cloud computing and web-scale computing require data center operations at unprecedented levels of scale and speed. This has created the need for new levels of abstraction, agility and automation. SDN arose from the need to automate networking at scale when cloud data centers started serving millions of users.

Unfortunately, the Internet of Things (IoT) is about billions of users. The transformation driven by SDN, such as network virtualization, was not developed with billions of users in mind. IoT will require mass-market levels of scale and simplicity. Automation will not be limited to the needs of a few sophisticated cloud architects and data center managers, but will need to address mass-market consumers.

The need for real-time array processing architectures for machine learning will drive high-bandwidth, non-blocking, full- or partial-mesh fabric architectures in networking. This will drive a need for higher level of abstraction than that available through overlay network virtualization.

Let's take a closer look at the networking requirements for IoT and how they go beyond what SDN provides.

First off, administration of the network is different for IoT. While SDN currently addresses automation and self-service through abstraction and automation, it assumes a fair degree of sophistication by its user. In contrast, IoT will require a much higher level of abstraction and simplicity. It is unrealistic to expect someone to update a Python script to get his self-driving car to pick a different route to work.

SDN is largely focused on providing automation and agility for large, centralized data centers. Similar to data caching, which optimizes storage location based on usage, IoT will require "application" caches running applications closer to the customer.

Business intelligence and analytics workloads are generally housed in large data centers running data mining technologies such as Hadoop in batch mode. The IoT, on the other hand, requires data centers closer to the user because of real-time needs. Smaller, distributed data centers will be necessary to provide capabilities such as offering real-time promotions based on customer location and activity.

Finally, as business intelligence (BI) evolves from reactive to predictive, computing styles will change. Clustered file systems and batch mode processing technologies such as Hadoop were sufficient for reactive BI analytics. Predictive analytics, however, require complex technologies such as machine learning.

Machine learning (ML) is a branch of artificial intelligence that deals with outcome prediction. It processes vast amounts of data to "learn." Most, if not all, of this data will be streaming real time from sensors. Imagine a jet engine with a hundred sensors streaming data real time to its airline, or other applications like remote health monitoring.

ML will require networks to connect streaming patterns of data in real time. Streaming data analytics, unlike traditional big data analytics, does not require centralized storage and retrieval. In many cases, the value of stored data actually diminishes with time.

One such network architecture is artificial neural networks (ANN). While in the past, ANNs have been used in niche applications, ML is likely to encourage wider adoption. Today's network architecture has for the most part transitioned from hierarchical connectivity to spine-leaf or CLOS fabrics. But the need for real-time stream processing, coupled with a desire to run applications closer to users, will drive full- or partial-mesh fabric architectures in networking.

Mesh-fabric architectures will call into question the benefit of overlay network virtualization, which is the leading manifestation for SDN. Overlay network virtualization implements tunnel overlays providing abstraction, automation and agility for micro segmentation on physical network fabrics.

The need for mesh fabric connectivity will cause overlays to merely replicate the underlay with no additional abstraction. Some other mechanism will be required to provide abstraction, automation and agility in mesh-fabric networks.