Kireeti Kompella, CTO of Juniper Networks’ Development and Innovation team, is pursuing a grand vision: To develop a network that is fully autonomous, that is able to monitor itself and configure itself to the demands of the applications it exists to serve, with no human intervention.
He calls this vision the ‘Self-Driving Network’ and sees many parallels between the evolution of computer networking towards this goal with the evolution of the automobile: from the primitive vehicles of late 1890s to today’s fully autonomous self-driving cars.
Specifically, he wants to see an industry initiative to create a networking grand challenge akin to DARPA’s Grand Challenges that did much to spur the development of self-driving cars from 2004.
“I really like how we made such dramatic progress in self-driving cars and I think something similar might be how we make dramatic progress in a self-driving network,” Kompella told Computerworld.
The US Defense Advanced Research Projects Agency (DARPA) staged its first Grand Challenge in 2004, offering a $US1m prize to the builders of any vehicle that could autonomously complete a prescribed 240km route in the Mojave Desert. There were more than 100 entries but none got beyond 12kms.
Several vehicles completed the 2005 challenge and in the third, 2007, challenge, over a 96km ‘urban’ course on a disused air force base, all finalists completed the course.
Networking grand challenge
Kompella said: “I want to issue a challenge that I think will be really valuable, that will change how we think about networking, and make possible things that are not possible today.” He added: “For us as an industry to make self-driving networks a reality vendors and network providers have to work co-operatively and vendors with each other.”
Kompella’s vision for a self-driving network is that it must be able to self-discover network elements, self-configure, self-monitor and self-correct, auto-detect customers, auto-provision to meet their requirements and to self-analyse and self-optimise.
For a networking equivalent of the DARPA Grand Challenges he proposes that a network must be able to “Run a data centre for six months with no human intervention (not even from afar) with no reduction or compromise in functionality.”
He told Computerworld: “I don’t think it would be right for Juniper to issue the grand challenge, but what the forum is I have not worked out yet.”
Cliff Grossner, senior research director, data centre cloud and SDN at analyst firm IHS, attended a briefing on the self-driving network at Juniper’s analyst conference in October. He said Juniper saw the concept as a necessary response to cope with evolution of networking.
“Juniper believes that continual advancement in the face of Moore’s Law slowing will require compute to grow by scaling out (becoming massively parallel) rather than scaling up, creating interconnected data centres, each with many 1,000s of servers required to handle applications such as augmented virtual reality, machine learning, IoT, and autonomous vehicles,” Grossner said.
“This will drive the creation of networks with 1,000s of nodes, too complex to be operated and maintained by humans typing at the command line interface. Automation will be essential—even imperative, according to Juniper—to change the current network management paradigm from one requiring continued human intervention to one where network management systems operate autonomously.”
Five key technologies
Kompella says five technologies are needed to make SDN a reality: telemetry, automation, declarative intent, local versus global views and decision-making, and he says most of these are in place.
“Telemetry will mean the boxes telling you what is going on, in a push format that is useable by machines. An automation framework enables the same thing to be done across multiple devices. Declarative intent is a description of intent where you specify at a very high level what you want done. This we implement today in our SDN controllers, Contrail and Northstar. The piece that is missing is decision making: this is where machine learning will come in.”
Grossner said he believed Contrail would give Juniper a good start towards a self-driving network. “It is still early days in delivering on this vision, and much development needs to be done on understanding the nuances of embedding artificial intelligence and machine learning into networking,” he said. “Juniper’s Contrail will provide an excellent perch point to introduce AI and ML creating a self-driving network.”
Machine learning will be key
Kompella said he had come to see machine learning as a key component of the self-driving network after discussing the idea with customers. “Machine learning is penetrating any field you can think off, and I believe we in networking are falling behind – we should be using the latest technologies.I have come to the conclusion that machine learning will be big part of this, and I went to a machine learning conferences for the first time in my life in Barcelona a few weeks ago.”
Grossner told Computerworld that Juniper’s acquisition of cloud operations management software company, AppFormix, announced on 1 December, would give the company’s machine learning efforts a significant boost.
Grossner said: “Part of the move to acquire AppFormix was to get the platform, which is based on AI techniques, but it was also as much about getting access to the people. I think it makes their self-driving network plans much more viable.”
AppFormix, according to Juniper “is redefining telemetry and operations management across software-defined infrastructures and application software layers.”
Self-defending before self-driving
Although commercial realisation of a self-driving network is a long way in the future, Kompella sees one of the first practical implementations of the idea as being a self-defending rather than a self-driving network.
“Security is where quick response and proactive action is really useful. If there is a security breach you need to know as soon as possible and you can’t do that without machine learning. This is the where the value of a self-driving network, or a self-defending network will be really high.”
Telemetry and visualisation
Grossner suggested the first steps towards the self-driving network would be in advancements in telemetry and visualisation dashboards, coming on the market in 2017. “I don’t think we will see any automated network tuning until 2018 at the earliest.”
He said Juniper was not alone in pursing this vision. “Cisco is moving into the analytics space with Tetration and Citrix has been looking at bringing AI and ML into tuning networks.”
Cisco announced Tetration Analytics in June 2016, with CEO Chuck Robbins describing it in a blog as “an open platform that provides data centre visibility at a level that’s never been achieved before, and at a scale never possible before.”
Tetration Analytics, he said "collects data in real time using software and hardware sensors … analyses the data with advanced analytics in real time, and delivers information critical for data centre operations by providing actionable insights in real time, through easy to understand visuals.”