Interview with Jayshree Ullal of Arista Networks on Network Virtualization and SDN
Matt and I had an opportunity to sit down recently with Jayshree Ullal, CEO of Arista Networks, and one of the luminaries of the networking world, to chat about her view of Arista, network virtualization and software-defined networking (SDN). Before Arista, Jayshree served as Senior Vice President of Data Center, Switching & Security Technology Group at Cisco Systems and was responsible for the Catalyst 6K/6500 and the Nexus line. This interview is part of a continuing series. If there are other network virtualization leaders, evangelists or technical visionaries you would like to see interviewed, drop us a note at firstname.lastname@example.org.
SDNCentral: Which core markets does Arista play in today?
Jayshree: Arista focuses today in the core datacenter and cloud networking markets, essentially anything in the 10G, 40G, 100G space. There’s a huge total available market (TAM) there, estimated to be $15B in 2016. We like large markets with large TAMs because small penetration can be meaningful to a company like Arista. Competitors don’t really get hurt with large TAMs since there’s enough to go around and additionally, we like it because our execution defines our success in such marketplaces.
When we started, the financials were a strong vertical for us, and 70-80% of our sales were for high-frequency trading applications. Today, that spread is much more even across multiple financial use-cases. In addition to financials, we also service cloud service providers and Web 2.0 companies, where we see deployments centered around networked storage, virtualization and high-performance compute. Across these verticals, what’s common is the requirement for high-performance transactions. With Arista’s high-performance low-latency networks, our customers can turn their applications into a business advantage, transforming a cost center into a profit center. At Arista, we focus on real operational and CapEx value instead of just speeds and feeds.
SDNCentral: Why is Arista making inroads against the incumbents?
Jayshree: Our success in these markets stems from our Extensible Operating System (EOS) that powers all our switches. When we first designed EOS back in 2004, the cloud was nowhere in sight. It took us 5 years to get to where we saw returns on our investment in EOS. Thankfully, we had the capital and unique financing structure to do it—no venture-capital funded company would have been given the leeway to take the route we took. During that time, we also managed to build up a significant software powerhouse by recruiting the right engineers—we have and will continue to seek out the top 5% of programmers.
EOS’s resilience, fault-isolation, open API platform differentiates us from Cisco and Juniper, and has been well-received by customers. Having said that, we don’t have all the enterprise features that Cisco has. After all, Cisco has had 15+ years to build up its products and we certainly do not have 15,000 engineers like they do. Nevertheless, our razor-sharp focus on building an outstanding foundation for our customers and their specific use-cases allows us to compete effectively. We did not set out with a swiss-army knife approach and will continue to maintain our focus on bringing value to our customers.
SDNCentral: Arista is viewed as one of the hottest startups to work for. What do you attribute this position to?
Jayshree: It’s our culture that drives our hot reputation. We strive to maintain a culture that is empowered at all levels of the organization. It’s a very open culture, neither top-down nor bottoms-up. We have a strong group of people here all with high IQs, but who don’t wear their IQs on their sleeves. When I first arrived, there were only 30 engineers, but the culture of openness had already been established. We’ve managed to retain that culture as we’ve scaled to multiple hundreds of employees—it is hard work though. What I particularly like is that the open culture transcends all parts of the organization, including sales, marketing and administration. I know it’s cliché, but A players do truly attract A players. We have been very disciplined about recruiting the top engineering talent, the top 5% in intellect. For us, it doesn’t matter if the talent is less experienced; we’re looking for employees with the mental horsepower. In fact, it’s interesting to note that when I first started, the average age at Arista was 40+ and today, the average age is 29. It shows how we’ve managed to recruit young talent here. We share a deep focus on datacenter products and solid networking protocols—that keeps us grounded.
SDNCentral: How does network virtualization and SDN tie into Arista’s core strategy?
Jayshree: For us, SDN isn’t just a single problem space, nor is there a single approach that qualifies as an SDN solution—we look at it multiple ways. What’s driving network virtualization is VM sprawl and the demands it creates on the network. Existing network vendors look at it the problem and say let’s just build larger layer 2 networks. Well, that’s when we run into the 4K VLAN limit and get stuck trying to vMotion VMs smoothly between those end-points.
We see a 3 phase approach to network virtualization:
- Phase 1. Server virtualization with vSwitch in the hypervisor. This was pioneered primarily by VMware. Phase 1 is a typical VMware deployment with vSwitches on the hypervisor. Interestingly enough, networking vendors talk about vSwitches a lot, especially in the context of companies declaring themselves to be the “VMware of Networking”. Well, my view is that anyone who proclaiming themselves to be the VMware of networking, is not. VMware is the VMware of networking. Who else has the penetration and density of virtual switch deployment that rivals VMware, or even comes close? The acquisition of Nicira by VMware strengthens my convictions.
- Phase 2. Going from virtual to physical. Once we have server virtualization in place, the next step is to bridge the physical and virtual networks. There are a couple of approaches here. One is to build your own switch, just like Cisco did with their Nexus 1000V. Arista took a different approach, and partnered with VMware instead. We didn’t try to replicate vCenter or replace the vSwitch already in place. We just mapped the vSwitch to Arista’s physical switch and provided tools to integrate with control infrastructure large customers already had, like vCenter. Many of these customers were excited with our offering, which includes tools like VM Tracer. These tools provided them greater visibility into their existing virtual networks and allowed them to troubleshoot their infrastructure better.
- Phase 3. Advent of VXLANs. I believe that VXLAN is as pivotal a discovery as the foundational IP layer 2 or layer 3 protocols. With VXLAN’s 16M segments, we are no longer limited by 4K VLANs boundaries. We can manage at layer 2 down to VMs and layer 3 up to the higher levels, with MAC and IP tunneling. Some critics don’t like VXLAN’s dependence on multicast for learning, but I don’t think this is an insurmountable problem—it can certainly be optimized and fixed in the future. And we’ve had multicast around for a long time and understand how it works, and its limitations.
VXLAN is a breakthrough for network virtualization that breaks down the walls between layer 2 and layer 3, physical and virtual. This allows virtualization to go a step beyond where we are today. VXLAN is basically SDN at work and already has strong multi-vendor support, from VMware, to Arista, to Cisco and Microsoft. VXLAN allows us to realize the potential of virtualization across servers, storage and the network.
SDNCentral: What network virtualization and SDN use cases are Arista delivering to customers today?
Jayshree: What we’re seeing primarily is VMotion across datacenters, both intra- and inter- datacenters movement of workloads. We expect to see cloud networks with 1,000s of nodes, scaling to 20K, 40K and beyond with VXLAN support. In fact, the nice thing about VXLAN is that many proprietary tagging methods can go away, as well as vendor-specific methods like QFabric. This is a big disruption to enterprises.
In addition, we’re also seeing the emerging use of load balancers to drive load steering across boundaries, with servers in multiple locations.
SDNCentral: You’ve spoken about SDN being multiple protocols and have also stated that many existing protocols can be applied to network virtualization and SDN. At the same time, Arista has also demonstrated OpenFlow running on your switch. What is your view of OpenFlow and what OpenFlow use cases do you think are interesting?
Jayshree: Fundamentally, we will support what customers want. We looked at OpenFlow 1.0 about 3 years ago in the summer and felt it was a good research project but we could not figure out the use case at that point. I must say that OpenFlow 1.1 was a good improvement and with versions 1.2 and 1.3, it is much better. Nevertheless, OpenFlow is not about replacing the Internet. Remember that OpenFlow is not SDN, but perhaps is the most hyped-up aspect of SDN. I’ve learned over the years that hype often is inversely proportional to revenue.
So our approach to OpenFlow is pragmatic, as education and research markets ask for it, we will deliver it and support it. Some of the emerging use cases for OpenFlow are wiretapping traffic (port-mirroring/spanning), particularly for the intelligence community. I also think using OpenFlow for isolation could also be interesting.
Perhaps the best thing about OpenFlow, or SDN (network virtualization), is that it’s brought back a new hope to networking. Networking is cool again. I can say that with pride because Arista was #1, beating consumer brands in LinkedIn’s top 10 most attractive startups in the Bay Area.
SDNCentral: Arista has provided innovation in the past like VM Tracer to interact closely with vCenter. However, will we see Arista entering the control and management market with your own solutions–perhaps extending CloudVision to go beyond configuration management and monitoring, into control?
Jayshree: We’re focused on manageability, not management. We make sure we have the right level of extensibility and expose the necessary northbound APIs to partners likeEMC, HP/Opsware, IBM Tivoli, Microsoft etc as well as to customers’ homegrown tools.
For us, CloudVision solved the single point of control whereby we’re running our dataplanes superfast at 10, 40 or 100Gbps, we need to have a way to connect our switches together to make them all look like one. We used XMPP (an efficient instant-messaging protocol) to enable that in CloudVision. It’s a natural integration point to other management platforms but we’re not looking to build our own.
SDNCentral: Some software companies are advocating ‘smart’ virtual switches coupled with fat, dumb physical pipes. What’s Arista’s view on this framework?
Jayshree: Well, there is certainly room for dumb pipes. For example, the Linksys switch I have at home is a dumb pipe, and many consumer markets will only require dumb pipes. Fundamentally, there’s a $20B TAM for switching. Some markets will require dumb pipes but there are many others, like today’s 10Gbps datacenters that require intelligence in their switches. It’s a case of different strokes for different folks and we certainly have a sizable market where we add value for cloud-scale data centers.
SDNCentral: What’s Arista’s position on L4-L7 technologies? Are there plans to build Arista into a Cisco Catalyst 6K replacement that leapfrogs the Cisco Nexus line?
Jayshree: Certainly at Arista we view ourselves as best of breed platform with superior software architecture. The Cisco Cat 6K was/is definitely a successful product line for the enterprise market that wanted a swiss-army knife approach, and the Cat 6K fulfilled that. Arista is not serving the traditional enterprise but key vertical segments looking for specialized applications. We are fulfilling the migration from enterprises to cloud and rather than branching into layers 4-7, we partner with leaders like F5.
SDNCentral: Do you think we’ll see another networking vendor like Arista anytime soon?
Jayshree: Getting Arista to where we are today took fortitude and vision, and required non-traditional financing. Venture capitalists are looking to put small amounts to get fast 20-100X returns, while we took a long term view with a $100M upfront investment. It’s a very different model that VCs will need fortitude to invest in, and so I believe that we’re unique in that respect.
SDNCentral: Thank you very much for your time, Jayshree, and we look forward to seeing Arista make further breakthroughs in switching and network virtualization in the near future!