Former Cisco data center chief Jayshree Ullal this week landed at switching start-up Arista Networks as its new CEO after leaving Cisco in May. Ullal spoke with Network World Managing Editor Jim Duffy about Arista's products, prospects and projections.
Former Cisco data center chief Jayshree Ullal this week landed at switching start-up Arista Networks as its new CEO after leaving Cisco in May. Arista, started by Sun co-founder Andy Bechtolsheim (who remains at Sun in a part-time role), made waves last year by packing 48 10-Gigabit Ethernet ports into 1RU for the data center. Ullal spoke with Network World Managing Editor Jim Duffy about Arista's products, prospects and projections.
What attracted you to Arista?
I looked over the summer at least three types of options -- clean tech and energy, software-as-a-service and video. I got approached for CEO and COO opportunities in bigger companies that I viewed as more directly competitive with Cisco. Quite honestly, my affections for Cisco don't change just because I leave the company. If you've been there 15 years, you don't want to just take the same thing and repeat it. Also, [I had a desire to] go back to my entrepreneurial roots. Cisco will always be the best big company I worked for. And I decided I want to go back to ground zero again.
What does Arista bring to the data center that's unique?
The price/performance [$400 per 10G port] is always a huge advantage. But what drew me was actually their software architecture. I worked with four or five different operating systems in my time, and we all talk about modularity of the software, but that to me is just the tip of the iceberg. The real core of the software architecture is understanding what problem you're trying to solve and designing ground-up to solve it. The pain point in the data center is that much of the networking software is very monolithic. You can bring in a kernel like BSD or Linux, but that by itself doesn't solve the problem. What was unique about [Arista's] EOS [Extensible Operation System] is separation of state from different processes. We have a transactional database where all of the state of networking variables is stored. Now you can do live patching, fault containment, live repair -- real-time, in-service software upgrades. So, it's kernel, state and then the specific processes, like virtual LANs, Layer 3 routing, or switching or [control line interface]. . . . This is not something you can just pull out of Linux or classical, traditional software. The potential of EOS . . . all in a 1RU platform is just very compelling. [Google] has been our inspiration in developing products and platforms because they represent one of the largest and most scalable data centers. If we design a Google-class data center, then we know we've covered the gamut all the way from small/medium to large.
Slideshow: What tech industry CEOs are saying about the economy
Can Arista go it alone?
History would show you that there haven't been too many companies that have successfully built a stand-alone entity in the last decade. I hope we can be a unique differentiation from history because of the innovation. We have over 1,000 man-years of networking expertise in the company, even though we're less than 50 people.
The customers are really looking for a specific solution for a specific vertical application. We're not trying to be the switch for all seasons, like I did at Cisco. We're trying to be focused on very specific applications for high-performance compute, Web, content delivery and vertical applications that really require cloud networking features . . . besides price/performance. Customers can use us as a complement to their existing Cisco infrastructure.
Do you plan to come out with a chassis-based core switch?
We definitely have a road map that goes beyond the 1RU. We feel, though, that the ability to pack 48 ports into such a form factor [is unique] compared to a lot of the modular platforms -- including ones I was involved with. In theory, if you rack up 1RUs, you can get 2,000 ports of 10 Gig E. There aren't a lot of modular systems today that can give you 2,000 ports -- more like 200 ports.
I think if we had done modular, it would be challenging, because in most environments we coexist with [Cisco] Catalyst 6500 or Nexus 7000 [switches] or a Juniper MX960.
Is the switch stackable right now, or is that on the road map?
No. We don't preclude it, but we don't have stackable software right now. People would like, over time, to see us have 40G and 100G between the switches as uplinks.
At some point, Intel will put [10G] on their motherboard. When they do that, the curve changes completely because right now, our switch port is cheaper than a host adapter. But as soon as that gets closer to $100 to $200, then you have the same phenomenon expressed in 1 Gigabit Ethernet -- a smooth, cost-effective transition to 10G. We expect that's going to happen in Q2 or mid-2009.
Will that take you into business workgroups and wiring closets?
When you're a start-up, it's very important to focus on what you do, but equally important is what you don't do. I will leave the traditional enterprise campus and wiring closet to my good friends in Cisco. We want to focus and be inspired by the requirements in the data center, and especially as it relates to this new trend of hosting clouds, either off the premises or on the premises.
Where is Cisco most vulnerable or challenged in the data center?
The challenge of anything in the data center, for Cisco and other companies, is always product transitions. They have a very large focus with the Catalyst, and now they have the Nexus. Positioning and appropriately deploying and achieving the right price/performance and functionality and design is a delicate task. None of these challenges are insurmountable. I also think they have some challenges in the area of partnerships. The partnerships with classical storage and server vendors are evolving for Cisco. How they strengthen and create differentiation with each partnership continues to be a challenge.