Arista 7170 Series Programmable Platform Hugh Holbrook and Nick McKeown

Hi, I'm Hugh Holbrook. I'm the vice president of software engineering at Arista, and I'm really excited to be here today to talk about our latest new platform, our switch. It's one of our densest platforms, a top-of-rack switch that has some incredible programmable capabilities in it, and we're excited to get it in the hands of customers because it's going to allow a whole mess of new possible functions and capabilities to be moved into a switch that previously required processing either on a host at slow speeds or in a dedicated smart NIC, which can be very expensive. So there's a lot of functionality — tunnel termination, traffic filtering, network address translation, deep packet inspection — that we're able to do. This is all possible because of some exciting software engineering and because of some fantastic technology inside the switch that makes this all possible. We're moving capabilities that used to be inside dedicated high-end routers or in network packet processors into switching through the use of some exciting technology from a company called Barefoot Networks. They've got a new switch chip called Tofino, which we're introducing in the 7170 today. And I've got here with me someone who I'm very excited to have with us. His name is Nick McKeown, someone who I've known for a long time since my days at Stanford, who's an industry luminary, really made some great advances, and also the founder of Barefoot Networks. So with that, I'd like to introduce Nick.

Yeah. So, Nick, could you tell me a little bit, just as we're starting out here, about Barefoot and what you provide and how we're using it, but kind of your mission and what Barefoot actually does?

Is that a programming language like C or C++? How does that compare?

P4 is a general-purpose language not tied to the Barefoot products, is that right?

So Tofino is a sort of different implementation of switching silicon or switch chip. There's fixed pipeline, fixed function, but I program the tables to modify it. And then there's the network processor kind of run-to-completion, where I write programs more like a program that would run on an x86 processor to process packets. And then there's Tofino, where I write a P4 program to describe the behavior.

And we see that with our engineers doing it, that they're able to make changes, understand it, and with less knowledge of all those gory details that require you to understand thousand-page register specs of silicon. We're able to achieve programmability in a way that's much more time-effective for our engineers, where we can work on the real problems rather than understanding those low-level details. But all of our customers don't write programs that run on EOS. What that programmability allows us, in conjunction with our customers, to do is to modify the behavior more quickly on behalf of our customers when we identify new use cases. So I think it's great for the people who have that appetite to go and define new forwarding models for the media entertainment industry, but they can also come to us with requirements and we can build that more quickly. So I think these are both interesting use cases in the evolution of programmable silicon.

Yeah, no, I think that's a real promise of the 7170 and the Barefoot silicon — that exact ability to have that conversation with the customer about what's possible and to do things that are not, as you said, standard networking. All the customers today, it's a competitive market out there. Everybody's looking for a competitive advantage, some way to make their business better, to perform their tasks better, to do more with the same silicon that everyone has, to provide some advantage. And being able to customize not just the protocols running on it and the control plane, as we've been able to do for quite some time with the OS, but being able to program the hardware — it's like another tool that the customers have. And that's what's so neat about this solution. To go to that next level. And the language is neat, and even if the customers aren't interested in being programmers of the networking silicon themselves, it allows us to react more quickly. And the fact that it's programmable, we can turn around those changes much more rapidly for the customers. And that's also very exciting.

I totally agree. I think the programmable silicon has really changed the landscape there of how fast protocol can evolve and how fast networking technology can evolve. I mean, the initial use cases that we have for the 7170, as you talked about, are things like custom tunneling formats inside a data center for data center gateway access, things that used to have to be done inside an FPGA or inside a host. And these use cases, we certainly didn't know about them — I don't think you knew about them at the time the chip was taped out. And we designed this all after the fact on top of the silicon, and that's just very powerful. Large-scale NAT — I'm not sure that you were thinking about large-scale NAT at the time that the chip was built, but the programmability has allowed us to build that capacity and solve a very targeted problem for a customer where this really makes a difference and allows them to come up with table scale that they didn't have before, combined with a performance that can only be achieved inside silicon. That previously had to be done either much more slowly or done at much higher cost by distributing it across servers or FPGAs.

Well, look, Nick, it's been super fun talking to you. It's always a pleasure. I really like talking about this. This is a great topic. It's been just a pleasure working with Barefoot over these last years coming up to the delivery of the 7170. I'm super excited about the product that we've got coming, about the partnership that we've got. And I just want to thank you for your time and for your help with us over these years. And I'm looking forward to delivering the 7170. It's available now, leveraging all of this cool programmable capability with the Barefoot Tofino inside, EOS running on top to provide Layer 2 and Layer 3 features that are common to all Arista products, and then packaging additional functionality on top of that to deliver targeted solutions for our customers, delivering new capabilities like telemetry, tunneling, scale network address translation, deep visibility and packet capture. And we're excited about possibilities that we haven't even thought of yet that our customers are going to come to us and ask us to build, and we'll be able to build because of the flexibility that we've got in this platform with the Barefoot silicon inside. So Nick, thank you very much. Wonderful to be working with you.