We have been talking a lot lately about the self-driving network. Last week, this topic was front-and-center at the AI Net Conference in Paris. Dr. Kireeti Kompella, Juniper’s CTO, gave a great talk titled, “AI for Network Automation and Service Design.” I also gave a similarly-themed talk titled, “Network Monitoring and Automation, and the Role of Machine Learning.” There were many parallels between the two presentations.
A key take-away from Kireeti’s talk was his definition of five stages of network maturity, culminating in a self-driving network. (Actually, there are six stages in all as the foundation is stage zero — the completely manually-operated network.) See the figure below.
The first step in the maturation cycle is simple automation with configuration scripts to configure routers using their APIs. This, of course, is not new. In fact, in the 1990s I developed software called IRRToolSet that configured BGP routing policies in the devices from high-level policy descriptions expressed in RPSL.
The second step on the maturity chart is monitoring of routing, traffic, performance data and other things. Step two provides visibility into the health of the network and facilitates capacity planning and security audits.
Step three is analytics, what I define as delivering actionable insights from the telemetry. Note that this is far more than just presenting the collected metrics and doing simple aggregations. Analytics may involve machine learning, event prediction, anomaly detection, trending, baselining, etc. Packet Design’s Explorer Suite is focused on steps two and three.
Step four uses these analytics to make changes to the network. Once the analytics are in place, writing automation applications is straightforward. I discussed several use cases in my presentation, and of course, Packet Design offers two commercial automation apps: the Explorer Path Provisioning and Traffic Engineering apps are examples of closed-loop automation processes.
Step five is the culmination of these technologies, where humans merely describe their intent and allow automation to take over the network’s operation. This involves AI as well as rule-based control. To me, this is the hardest step – but we are working on it at Packet Design!
On his chart, Kireeti asserts that the industry as a whole is at step one (and actually, closer to step zero). This surprised me. I asked whether he thought some network operators are at step three. He acknowledged that some are; of course, he was referring to Packet Design customers. We agree, our customers are at least two steps ahead of the industry!
To illustrate this, one of the automation examples Kireeti described was illustrated earlier in the week at our Customer Symposium. Ron Ripley of Shaw Communications presented how thousands of business customers were safely migrated from older routers to newer ones. Ron explained that they were able to complete the project in just a few weeks with analytics-driven automation, using the Explorer Suite’s REST APIs, instead of six to twelve months that it would have taken doing the work manually.
While fully self-driving networks are still in the future, that future is not so distant. Kireeti’s chart is an elegant tool for assessing a network operator’s current state and provides guidance for moving through the maturity steps.
Where are you on your journey to a self-driving network?