Dr. Raj Yavatkar, Ph.D. - Chief Technology Officer, Juniper Networks - Innovations In Networking

Sure. I worked with Professor David Comer, who was one of the members of the so-called Internet Architecture Board, which basically included companies such as defense contractors like BBN, Sun Microsystems, and AT&T, and so on. One of the projects I worked on as a research assistant was called Cypress Network. It was the first nationwide network created in academia before NSFNET, which used 9600 baud lines with simple workstations as packet switches to connect the network across the country, starting from Boston all the way to California and the Midwest, and so on. Out of that came the thesis idea: how would the evolution of the network happen, and what is required to show that you can scale the internet speed? Of course, as you said, now we are at a very different speed, and at that time it was a very different era, but the idea was to define the architecture for packet switch networks along three dimensions. How will each packet switch scale to high speed using multiprocessors, using all software, implementing nothing in hardware? Second, as part of my thesis, I'm proud to say that I was one of the people who introduced the term 'flow,' which is used very commonly today. It was introduced in a paper I published in '89 in Infocom. The flow basically says that rather than looking at packets going from one machine to another, look at the applications and application traffic, identified by these five things. If you're doing video streaming, you're going to have different characteristics in that flow versus sending some audio traffic or just a file transfer, right? Part of the thesis was also to look at what sort of congestion management and flow management techniques you'll need in the network so you can guarantee certain quality of service for audio traffic, which requires low latency. You want to deliver audio packets in 250 milliseconds or so before you start noticing gaps in the conversation. Video, you want bigger flows delivered, and sometimes delivering out of order or losing some packets is okay. A file transfer is another example where you need to do large file transfers in a reliable way. Looking back, some of the things are still relevant, but some of the things have evolved way faster than we imagined.

I've been very fortunate to work with really smart people in my career. When I was at Intel, I used to work with Professor Andrew Campbell who was at Columbia University and his students. One of his students came to work with me as a summer intern, and I hired him after his PhD, Michael Kavis. I would say he's more of a Greek mathematician than an engineer, very smart guy. This particular paper, he's the first author, it was a joint work we did. At that time, video processing was in vogue, with a lot of surveillance cameras coming into the picture to address security issues. But if you look at the video stream coming in at an airport or a shopping mall, there's lots of motion happening, but most of it is background. Something new happening, to be able to detect those fast while processing so much video information, is very hard. The idea here was to only focus on tiles of the video, tile it into different small images, and only look for, based on time series analysis, what has changed and remains changed for a long time, so you can focus on those. That becomes very fast, and we were able to show that it scales when you're getting multiple video streams in a very busy intersection or airport. That was the work, but I give Michael the credit because he was the real mathematician behind it.

In the late 90s, the dot-com boom created the internet being used for the first time in consumer commercial ways because of HTTP, web browsers, and all that. The amount of traffic in the network was literally doubling every six months, and using traditional computer processors like Intel-based processors to do routing was not sustainable. The idea Juniper came up with was to have a custom chip designed for packet processing. The first router, the M40, provided throughput an order of magnitude higher than what was possible at that time. It followed up with creating a series of products, and that's how the company grew. We continued that path and also got into enterprise networking and so on. The biggest change that happened ahead of the industry recently was in 2019 when Juniper acquired a company called Mist AI. Mist AI was applying a very different approach to managing wireless networks. They said we are just going to have access points you see all over in the building, anytime you use a Wi-Fi network on premises, they will be connected to the cloud, everything will be managed from the cloud, and we collect data from the network continuously. That data will be fed into machine learning algorithms to do automated root cause diagnostics, being able to tell that a problem exists before even the user notices. Now we have about seven years of data from the network collected, trained models which we can apply to wireless networks, campus branch networks, access networks, data center networks, and WAN routing. That's the big advantage we have, and that's what led to the acquisition by HP Enterprise of Juniper for networking differentiation.

Sure. Let me start with a simple example. You are on an iPhone or iPad connected to a Wi-Fi network, and many times people start seeing slower performance or a glitch on a Zoom call. The traditional method is you file a troubleshooting ticket if you're in a corporate environment, then somebody from IT looks at it, contacts you, tries to find the problem, and fixes it. Ours is different. All that data is going to the cloud, and based on that data and all the other data being collected from other devices as well as in the network itself, you start correlating. Machine learning is very good at taking large amounts of data and finding patterns easily and automatically. You can train models to do that. So if you train the model, now when my iPad starts seeing a glitch, it already finds out what the problem is. Maybe the switch inside the network to which I'm connected is having a problem, running out of buffer space, or packets are getting overrun. If I know the reason, I can start taking automated action. The self-driving network is full of sometimes tens of devices, hundreds of devices, all interconnected to let traffic go from one endpoint like my iPhone to a cloud where I might be using Gmail or Google Drive. If I have data telemetry collected and I use this constantly to do correlation with machine learning algorithms, then I am able to find anomalies, find the root cause, and fix them in an automated manner. This has happened publicly. Some of our customers, Fortune 10 companies, have publicly stated that they have reduced their trouble tickets by 90% using our self-driving network.

The self-driving network is an evolving concept; we are going to learn continuously. When I step back and look at this from my role as CTO, we came up with this concept of pathfinding innovation. It basically says how can we continuously create a system by which multiple people can come up with ideas? That's a funnel of ideas. We funnel them into some kind of a pathfinding effort where we can quickly try out the ideas, see which ones are likely to lead to good results or get good impact, and then incubate new software products to add to the current portfolio. That's the pathfinding concept, and we have been driving it. Ideas come from many places. We have four pillars of innovation. The first is very much the self-driving networks you mentioned. The second is sustainability, because we have to be good corporate citizens, but not only responsible, it makes business sense to be sustainable in our practices and how our products are used by customers. Third is quantum networking. The world is going to go to quantum computing and quantum networking. Some people say it will happen in 10 years, some say 5 years. We have to find what is real and what can be used. Lastly, 5G. The internet has evolved; no longer is it about sitting in the middle of some big network. All of the traffic is mostly driven by mobile internet, so 5G and its evolution is very important. Those are the four pillars. Let me go through each one. For self-driving networks, we want to provide assured service experience, which means you not only get connectivity but also assured service experience with respect to performance, security, latency, throughput, and so on. We have been evolving the machine learning models to collect data from everywhere in the network, not just the network but applications, and start correlating across all of these parts in the application stack to apply machine learning to provide guaranteed service experience. Second is sustainability. As we do products, customers come from different sectors: enterprise, service providers like AT&T, cloud customers like Amazon. When you over-engineer these products and deploy them, they use more energy than necessary because only a subset of features are being used. We came up with a mechanism where, in a running network for a particular customer, we can turn off features and parts of the devices to save power. Secondly, customers also over-engineer their networks to be ready for peak capacity. For example, if you're a network provider to a stadium and suddenly the 5G network usage spikes, we can turn off parts of switches so rather than using four different paths, you can use one path because utilization is only 10%. That's a sustainable network. The third pillar is quantum computing and networking. One of the biggest problems in the network is when you use security like IPsec tunnels. People use VPNs, and we do symmetric key exchange where both parties have to agree on the keys to encrypt traffic. That is prone to attacks. If quantum computers get more powerful, they can break those keys. We need to protect the connection using an out-of-band quantum channel. There are startups that have come up with quantum channels, so we worked with them and now have a product for IP VPN tunnels that does quantum key distribution to protect against key breaking when quantum computers become real. Quantum computing is becoming real slowly; people on Wall Street are using some financial services quantum computers with probably a thousand qubits or so. It's important to start looking at this evolution and not wait for it to disrupt our networking industry. Lastly, 5G. There's a new initiative called Open RAN Alliance, Open Radio Access Network. We developed what we call the operating system for that, which is completely open. Just like you can use Windows or Linux, anybody can write applications on top of Windows, and it can run on any platform. Similarly, our RIC can run as an operating system for the network, supporting radios from any vendor, and third parties can write applications to optimize the network, make it more energy efficient, more secure, and provide guaranteed service for different slices.

The kind of investment that companies used to make in long-term research is no longer practical. A lot of companies are not able to do that. At the same time, there's a need to look ahead at where the industry will go and still be able to invest in it. Juniper Beyond Labs is an initiative we started. It's an umbrella across the company to see how we can invest in innovation and pathfinding that is still practical and will lead to practical things. We can make some bets in a way that we can early on decide whether those bets are going to pan out or not and decide whether to invest more or not. That's where this idea of pathfinding and incubation comes in. You start with a big funnel of ideas, start narrowing them down, and we have teams across the company in the areas I mentioned, those four pillars. That's the kind of investment we can afford. We also started a university research program. We fund university research because that way there are no strings attached; they develop their own research but collaborate with us, so those ideas can come in. We can hire those engineers who graduate with Masters or PhDs. The third effort is partnerships and collaboration with other companies like IBM, Red Hat, and Intel. We use a term called 'minimum viable demo,' not a minimum viable product. Typically, a startup develops a minimum viable product they can start selling to customers. The minimum viable demo concept I came up with is similar in that it's a demo, not a product, but it's not a science fair demo or a throwaway demo. It's a thing we build that we can hand over to customers. It's robust enough for customers to try and give us feedback. It's neither an MVP nor a demo you just do at a kiosk at a conference; it's somewhere in between.

That is the right question. Let me start with the second one and go to the first, because the second one I'm very passionate about. I do believe that the system of innovation that the US has created is long-lasting for the following reason: the startup ecosystem and the concept of venture funding fuels innovation, even in universities and academia. I used to be a professor a long time ago. At that time, professors gave away whatever they did; it went into the open source movement or different kinds of open systems. But now, professors even in the middle of the country think about innovations and how to commercialize them because they are investing in innovation. Universities get funding from funding agencies, but they are also looking to take it to the market. I think that sort of close feedback loop, where you have some initial funding from the government or industry to do research in universities, which leads to venture funding and innovations coming out, is a great way to continue to innovate. What is important is that the government continues to fund research in academia through the National Science Foundation, DARPA, Homeland Security, and different mechanisms. That has to continue because that's a seed. We also create internal mechanisms. For example, we have a Juniper annual conference internally run like a research conference in academia. People are required to submit abstracts of papers, posters, or demos of new work they're doing. This could be two or three engineers getting together and doing something outside of their main job or part of their main job. The conference becomes a way to exchange ideas and demonstrate things. We give prizes and rewards, and connect them to the right marketing or product management resources so those ideas can be taken further. The second thing is a very strong mentorship program. Inside, we have something called Juniper Gigs. People can work on it during the weekend or free time, whatever they want, two or three people. They advertise the gigs, somebody matches with the gig based on their interest. Suppose they are today working on platform validation but they are looking at all the generative AI and OpenAI and think they're missing out in their career because they're not working on those things. The gig is a way to give them an opportunity to work on new things. I might post a gig saying, 'Can you apply generative AI to create better demand generation emails for marketing?' Somebody signs up for that gig, has to learn new tools from OpenAI or other open source models, and apply that as part of doing the gig.

I always tell people not to watch predictions, including mine, because they're more often wrong than right. We always get surprised by the end of the year where the year ended up. You have to be humble about these predictions; you think you have a crystal ball but you don't. What I've seen in terms of changes by the middle of this year is that generative AI is just moving on very fast, and nobody can doubt that. But one of the things I'm also noticing that people are not paying attention to is optics. The network is rapidly evolving in terms of the amount of bandwidth needed to support these machine learning things. Innovations are happening in the optical domain, including optical switching. The second part is that we always underestimate the impact on security of new technologies. I'm beginning to see that generative AI is not just about deep fakes and impacting elections; it's generating more kinds of attacks, attack vectors that we can't even imagine. They are completely generated by machine learning, able to explore the search space for attack vectors in a way that human beings cannot. That's going to be a bigger threat than we anticipated.

Thank you for having me.