How A Rocket Took Off: Juniper Founder Pradeep Sindhu (Part 4)

SM: This was right at the time when Internet was beginning to emerge in mainstream communities.

PS: Exactly. Netscape had just released its browser. I had the opportunity to see the number of Web sites and the amount of bandwidth being used. It was absolutely exploding. It was doubling every six to nine months. I have always been a big believer in the idea that when you see exponential growth, you had better look underneath. In nature, you do not see exponential phenomena last very long because you run out of resources. Here was an example of bandwidth doubling very often. My analysis was that there was a latent need for any-to-any worldwide connectivity, and even the protocols had existed for a while. Things like FTP and telnet had existed for a while but needed experts to use them. Email had been around since 1976.

The browser allowed ordinary people to make use of the Internet. The combination of the World Wide Web and the Internet browser removed the barrier. Now you saw how useful things could be. This meant that bandwidth would keep exploding, so I had to find where the next limits would be. The router seemed obvious.

I found that Cisco, Bay Networks, and 3 Com all built routers by programming an embedded computer system to move packets. The CPU was being used for packet forwarding as well as to run the routing protocols. The technology being used to build these things was five to six years behind the technology that I had just finished working on. The opportunity in this space seemed clear to me.

SM: Was your thought to use different processors for different functions?

PS: No, the thought was radically different. It was not to change the definition of what a router is but to change the implementation. Let’s suppose we are going to do IP forwarding. It was clear to me that the router needed to operate on two planes. One was the data plane and the other was the control plane. The control plane was where the routing protocols ran, where you did the configuration of the machine, and it could only benefit from general-purpose processors. I was not going to add anything to that problem, so we chose x86 as the engine. In retrospect, that turned out to be a very good decision.

The second part is the data plane. That is where most of the packets move. The key was to be able to do this at very high speeds. We needed a new architecture. It was not a problem of caching. It is a very complicated problem. You have to do a bunch of things very, very well. You need to parse packets as they come in; you need to buffer packets, prioritize packets, switch packets between ports at very high speed;, and you need to interconnect all packet forwarding engines together into a single system.

At that time, nobody had taken an entire IP packet stack and built it in silicon. In our first product we put IPv4, IPv6, and MPLS to build a specialized router for the core of the network. The capacity of the first machine that we built was CO48, which is 2.5 GB per second. Today that is a toy, but at that time it was pushing the limits of what was possible.

SM: This concept all came about in the summer of 1995 while you were in France?

PS: The idea of working in the wide area network occurred in France. From September to November 1995, I worked on a design. They key problem then became money. I had left PARC and was doing this completely on my own. The three VCs I wanted to talk to said that there was a large existing player which took away the scope for innovation. I then talked with Vinod Khosla at [Kleiner Perkins Caufield & Byers]. In December 1995, he decided that he liked the idea and took a chance on the company.

SM: How long did it take you to bring the product to market after you got funding?

PS: The first thing I did was hire Dennis Ferguson and Bjorn Liencres. Dennis has a tremendous background in networking, and I had worked with Bjorn at Sun. One represented the software side and the other the hardware side. I was the first CEO of the company. We started the design in earnest in January 1996. Almost all of the people we later hired were people who we knew and had worked with before.

In terms of design validation as to what to build, we relied primarily on Dennis. He had built a router before and had worked for service providers. He understood the problem from both sides. We had told VCs that this was not something that we could pull off in a year. It required fundamental innovation, chips to be built, and software to be developed from scratch. We figured that it would take us a good two years. It ended up taking two years and three months.

This segment is part 4 in the series : How A Rocket Took Off: Juniper Founder Pradeep Sindhu
1 2 3 4 5 6 7