Arista Networks Inc. is being recognized with this month's Network Innovation Award for its Universal Spine, which relies on a distributed leaf-spine approach, rather than massive core routers, to serve the needs of large enterprises and internet service providers.
Universal Spine is based on Arista's 7500R Series of high-performance switches and routers, the most powerful of which has a maximum capacity of 15 Pbps.
SearchNetworking spoke to Martin Hull, Arista's senior director of systems engineering and platforms, to find out more about what the vendor is doing with the Universal Spine platform.
This interview has been edited for length and clarity.
It's been a little more than a year since you announced Arista Universal Spine. How has it been received in the marketplace?
Martin Hull: The product is the culmination of a continuous evolution of platforms we've had in the market since 2008. But in terms of the Universal Spine, that was the bringing together of additional capabilities afforded by improvements in the underlying silicon, the software evolution of EOS [Arista's Extensible Operating System], and then the market adoption of traditional switching platforms in roles that had previously been reserved for routers. We've had good success. I don't know that we necessarily had goals or targets, but it has certainly met our internal expectations in terms of the types of customers and the volumes.
What did customers want to know about using Universal Spine in their operations?
Hull: We've been deploying leaf-spine -- two-tiered -- networks for many years. So, this was the concept that the spine could actually be used as the next tier up. Many customers have been using [this] class of products in their data center networks for many years. And so there was already an implicit acceptance that the product was fit for this purpose.
[Universal Spine] was about enhancing the products with additional density, additional performance and additional scale. We thought it was the right product, the right time and the right set of customers.
OK, so let's talk a little bit then about some of the problems and challenges your customers were expressing to you and how the Arista Universal Spine addressed them.
Hull: So, the core classic issues are density and price. To get a device with the capacity that customers required meant a large system with, generally speaking, legacy proprietary technologies that [customers] didn't necessarily want [and systems] with high price points and high power-per-port requirements.
In traditional router platforms, density is significantly lower, so price and power per port is significantly higher. We were able to address those issues with Universal Spine, offering higher density, higher capacity and higher throughput, while also reducing the price and the power required per port.
What are some of the forces that allowed you to drive per-port prices lower?
Hull: At one end of the spectrum, we're riding Moore's law, in that every 24 months, you get higher-capacity silicon for effectively the same power budget or the same price. What we've also seen is that the chips themselves have become more feature-rich. That means they can now be used to replace custom silicon from routing companies where, you know, the whole bunch of legacy proprietary features had been baked in.
What we've seen is that the customers have largely decided they don't want or require the large set of legacy features. What they want is open IP-based networks. So, the merchant silicon has become more capable; the customers have moved their requirements to the point where they've simplified their environments. So, there's been a coming together of those two things. So, that's one of the ways in which we're able to address this as a price performance curve -- riding Moore's law -- but also the customers themselves recognizing that merchant silicon offers all of the capabilities they need.
With all the attention being paid to making networks more software-based, how does Arista Universal Spine and the Arista leaf-spine architecture dovetail with those developments?
Hull: The EOS operating system that runs on the Universal Spine provides customers with open API access. A number of customers today are using this capability in their private backbone networks, and they're using their own internally developed controllers to program our devices. Others have software-defined networks running, and they programmatically manage and provision, control and steer traffic directly onto our Universal Spines.
Let's talk about some of the upgrades you've made to Arista Universal Spine. What does FlexRoute do, and how does that help customers?
Hull: FlexRoute is how we can help meet scale requirements. FlexRoute is the technology we developed that allows you to [perform functions, and at scale] on a chip [that the chip itself] doesn't natively handle. It's software and hardware working together to get more out of the same.
So, how does it work within the Arista leaf-spine architecture?
Hull: As we look at the use cases for routers connecting directly to the internet, there's an accepted scalar number [that states] the device needs to hold at least 1 million routes. That's roughly the size of the internet table with headroom -- room to grow. FlexRoute is how we take standard merchant silicon and marry that with our EOS software to get over 1 million routes programmed into the hardware, without playing any additional tricks or only loading partial tables.
How are you engineering Universal Spine and FlexRoute to handle emerging demands from your customers?
Hull: We've worked over the last year to enhance FlexRoute. The 7500R Series now handle in excess of 2 million routes now. The internet obviously hasn't doubled in size in the past year, but what we're seeing is customers asking us to hold multiple copies of the forwarding table, and to program multiple copies of the forwarding table into line cards.
There's also been adoption of the Universal Spine in some NFV [network functions virtualization] environments where we're partnering with companies like Affirmed Technologies, [which provides next-gen mobile services]. Another area is internet exchanges. These have been, traditionally, flat Layer 2 environments, and you essentially buy a circuit and that allows you to connect to everyone else in the exchange. What we've seen is a move by these internet exchange points [IXP] to deploy Ethernet VPN so they can have a more intelligent Layer 2 network. So, we've enhanced EOS to support those efforts and help improve the value of IXPs.
Tracking the evolution of data-class switches
The drive toward open APIs is changing networking
Higher speeds driving the data center