Reprint

THE CONNECTED ENTERPRISE February 22, 2013 How got out in front of the SDN craze Arista CEO Jayshree Ullal says ‘cloud networking leader’ complements Cisco

By John Gallant, Network World Arista was able to take advantage of that disruption in hardware. Our top five oday, the buzz in networking is The second is software. We were very differentiators are all all around software-defined net- inspired by Cisco’s software focus on the works — and nothing could make enterprise side, Juniper’s on the service tied to our software.” Arista Networks CEO Jayshree provider side, and we saw that we could — Jayshree Ullal, CEO, Ullal happier. Ullal spent 15 years build a purpose-built, modern operating Arista Networks Tat Cisco, where she ran the network giant’s system only for the and the core switching and data center businesses, cloud. We didn’t try to do it for general- before joining Arista, which was founded purpose networking. We really focused on as a cost center, but really build it as a by co-founder and Chief our mission, which is high-performance profit center by addressing the applica- System Architect and applications for the data center and cloud. tions themselves. We early on entered the , a pro- It’s called Extensible high-frequency trading market to under- fessor of computer science and electrical (EOS) and there is no networking operating stand their trading algorithms, map it to engineering (and fellow Cisco alumnus). system that is as modern, self-healing and the latency requirements. That became an Ullal says Arista’s data center switches were resilient, and [designed for the cloud]. instance of a high-performance financial born to support SDN and provide both the And the third, speaking of that, is the cloud where they started building the net- power and flexibility required for today’s cloud itself. The enterprise market is work for that application separate from the highly virtualized corporate and cloud data shifting. Every CIO is being demanded enterprise network. centers. In this installment of the IDG Enter- a strategy on what they are doing with In Silicon Valley, a large number of Web prise CEO Interview Series, Ullal spoke with the cloud in terms of applications and 2.0 providers, whether they’re search en- Chief Content Officer John Gallant about infrastructure. Whether it’s a private gines or social networking, the kind of scale the reality and hype around SDN, and why cloud, a public cloud or a hybrid cloud, they build is just unbelievable. It’s 100,000 the data center requires a different network these are becoming an important piece of nodes, and increasingly, one machine, one than your father’s general-purpose Cisco the strategy. As Amazon innovated on the physical server, is not one node. That’s 20 net. She also explored how her work at Cisco application side, you can think of Arista as virtual machines, which means you could shaped Arista’s strategy, and shared insights really providing that market disruption on be enabling 100,000 physical nodes but on how Arista’s partnerships with VMware the networking side. you are really enabling 1 million virtual and Cloudera are making it easier to move to nodes. There’s huge virtual machine sprawl cloud and embrace big data, respectively. Explain the cloud angle in a little and physical sprawl. The CPU at one point more depth. What were you set- wasn’t being fully utilized. But now, with There are a lot of networking al- ting out to do to support or enable the new multi-core CPUs, the pressure is ternatives out there. Why should cloud? back on the network. That’s why whether someone buy from Arista? More and more people are outsourc- it’s a private or a public cloud, the Web 2.0 Arista saw three disruptions in the ing to modern applications — whether it’s companies are moving massively to high- market: a hardware disruption; a software Salesforce.com or Amazon itself. [They’re density 10G, 40G and 100G [networks] disruption; and a customer buying disrup- supporting] high-performance computing, that are requiring a new type of architec- tion, which in my mind is the most impor- or high-frequency trading or, increasingly ture and new software as well. tant thing. You can invent all you want on now, big data and network . the technology side, but you have to see the The network infrastructure needs to adapt. What are the things that make you customers changing their market position. It cannot be so monolithic. It cannot be one different than a general-purpose The hardware technology disruption was physical port equals one VLAN equals one networking company like Cisco? that in the 1990s, the only way to build any . It really needs to be much At the highest level I would say our kind of high-speed networking was through more massive in scale. A typical enter- software, our EOS. It’s open, it’s built your own in-house ASICs [application-spe- prise network is a 10,000-node, three-tier out of straight . But then we added cific integrated circuits] and specialty chips. network, and we were able to build a much what we call multi-processing, state- That’s not true anymore. We have from three flatter, fatter topology at Layer 2 and 3, us- oriented software that allows you to do the to five vendors available, whether it’s Intel, ing what we call the leaf-spine architecture kind of things that you could only do in Broadcom or others, supplying us much of that can scale to 50,000 to 100,000 nodes. mainframes and servers. It’s funny how the silicon. They are sometimes an order That was our first premise. hardware changes every 18 months in of magnitude better in power, footprint, The second [thing we focused on] was networking, but software doesn’t change density, latency, and performance and scale. application delays. Don’t build a network for decades and has remained monolithic for so long. Our top five differentiators are problem. We call this “from A to Z analysis.” er guys really like it and other times it’s high- all tied to our software. We can do automation, we can do zero-touch density 10GB. Another application is big The first is that we build, without us- provisioning, we can do a suite of functions data. Storage is no longer just a fibre-channel ing any proprietary components, active/ here because data is coming at such amazing SAN — you will start needing 10GB storage active networks that can scale to 50,000 speeds, structured and unstructured, how for iSCSI or more and more Hadoop clusters and 100,000 nodes. Other companies try do you sort out what’s relevant and how do with direct-attached storage. That becomes to do that with proprietary technologies. you monitor, how do you tap, how do you do another very interesting Arista project. You may be aware of Juniper’s QFabric real-time captures at 10 gigabits and terabits Virtualization, the VM sprawl. Another one or Cisco’s FabricPath and OTV [Overlay when the data is moving so fast? We’re not we’re starting to see more of is huge media Transport Virtualization]. We are able to just building enterprise features. Cisco’s rendering, and video applications that are do it in a standards-based fashion, and ev- done that really well for the last two decades, pushing the envelope of bandwidth. Where ery one of our networks interoperates with that’s their market. But yet if you look at the the application intersects the network is the Cisco routers, Juniper switches, NetScreen way servers are sold today, only half of them common theme through all the projects. firewalls, you name it. are going into an enterprise application. On the other hand, Arista has to walk be- The second is, because of the software, The other half, which are high-performance fore it runs. We’ve been growing at the rate we were able to bring to the data center and computing and Web, are going into the cloud of one new customer a day since we started cloud what we call self-healing resilience. applications. They don’t require traditional shipping. We now have 1,700 customers. Usually, redundancy and resilience means enterprise features. Just like mainframes Deployments usually start small, then they buy two of everything and connect them moved to client server, enterprises are mov- get really fascinated and intrigued and in case one fails. It’s great for the vendor to ing to more HPC and Web, and those fea- appreciative of EOS, and all of its opera- get two of everything. But we were able to tures are much more about reducing OPEX tional advantages, how open it is, how easy do it right in our software. Today, you look and improving the orchestration and traffic it is to use. The training is very easy and at software agents and how they interact. If visibility and data analysis. a Cisco CCIE expert would be able to use you have a memory leak in software today, The fifth and final differentiator is Arista right away, because we have similar and the agents talk to each other in a tradi- network visualization. What VMware did command-line interfaces and operational tional , they do to servers with server virtualization, we be- look and feel. Where we don’t have to so with something called IPC, inter- lieve jointly working with VMware we can invent, we don’t. Where we had to invent communication. But think of the cloud do with network virtualization. VM sprawl for these specific use cases we do, so most where you have, like we described, 100,000 has created network sprawl. Arista and often it’s a use case or a project. Sometimes of these, the multiplier effect of failure is VMware, together with a number of other it’s a data center build-up. After they use huge with this inter-process communica- vendors, Broadcom, Cisco, etc., defined to us in one project, they’ll say they want to tion. Arista chose a publish/subscribe mod- me what is one of the most breakthrough consolidate data centers. I would say 10% el using a built-in SYSDB database, where specifications in our industry — VXLAN, to 20% of them are now standardizing on the state of every software agent is stored. virtual extended LAN. The VLAN, as a unit, Arista as their data center strategy. Because that’s not human-generated, it’s the is something we all grew up with and in- most resilient piece of code. Let’s say you vented back in the ‘90s. It’s been with us 25 You mentioned about 1,700 cus- have a failure. We automatically track the years, way too long. VLAN boundaries have tomers. Give me a sense of your failure and contain it. Then we repair it. We plagued the deployment of virtualization business progress to date. actually up a new agent. Today’s enter- because you’re limited to 6,000 VLANs or We’re not supposed to [talk revenue] prise agent manager has no maintenance 16,000 VLANs, and you’ve got many more but the company is very young, it’s only windows. So they don’t have to know. virtual machines. So therefore, you’ve had a 5 years old. We’ve gone from 30 employ- The third [differentiator] is that we are vi-admin manage one, the virtual network, ees when we started to more than 500. I open and programmable. You hear a lot of and the command line interface or Cisco guess the biggest thing I’d leave you with talk about SDN these days, and one has to admin manage the physical network. These is that in the beginning we were a market separate the hype from the reality. The es- two worlds need to come together. Arista, leader for financials and high-frequency sence of SDN to me is, first of all, build open working particularly closely with VMware, trading which, as you know, is a tough interfaces and allow your customers to write has been able to bridge that gap between customer. We’ve always had to go into to their applications through our at the network physical and network virtual, us- mission-critical [environments] and we northbound level, and at the southbound ing VXLAN. VXLAN all of a sudden opens didn’t have it easy ever. It wasn’t like we level our devices must be programmable. We up the boundary from 16,000 to 16 million were in a little lab somewhere. We believe didn’t call it SDN back when we developed possible entries. So we’re very excited with we are today 70% to 80% market leaders this, we called it EOS. The extensible in EOS the technology we demonstrated at [the in high-frequency trading. In 2008, 2009 is [in reference to the operating system be- VMworld conference]. and even part of 2010, that was 70% of our ing] very programmable. Every aspect of our business. Today it’s diversified nicely into software, whether it’s at the hardware plane, Is it deployed now in the market? three areas. The first is financials. The sec- at the device plane or the software plane, can Very early. We are one of the first to come ond is what I call Web 2.0, and the massive be programmed. That’s a huge advantage. out with it. We showed it August 2012, and scale of their deployments, the cloud scale, We find ourselves in a fortunate position we showed interoperability with VMware, really. The third is cloud and service pro- that as the SDN market is evolving, our EMC and F5. We shipped a product based viders. Every service provider is looking to network is already open and programmable on it, the Arista 7150, in November. be a cloud vendor. In all of these three cases and SDN-ready. we are being looked at as the innovative The fourth one is big data analysis. Data Say I’m a big Cisco installation alternative to traditional legacy players. analysis and traffic visibility is becoming today. When would I talk to Arista? a real weakness, because, as you know, What’s the need that opens the You were at Cisco for a long time. we can all talk about improving price, door? Yeah, 15 years. I intended to be there two performance and CAPEX, but the biggest It could be project-based or it could be a years. But I was there 15 years, two years at cost center in networks is OPEX. There strategy. When it’s project-based, it’s usu- a time. are three ways to solve OPEX issues: Stop ally that you’re deploying high-frequency buying gear, outsource your gear or make trading or you need a high-performance So how did your experience at your technology do better work. We believe compute solution, usually InfiniBand and Cisco shape this? technology to solve the problem is far better get reviewed. Sometimes Infini- I had a big hand in shaping Cisco’s enter- than outsourcing or throwing people at the Band gets chosen because the supercomput- prise switching strategy, and it helped me

2 february 22, 2013 www.networkworld.com appreciate what to do and what not to do. you replicate it, it’s a huge multiplier. The Let’s go back to SDN. What’s real At some level I don’t feel we compete with third area I’d say we have really evolved and what’s false about SDN? Cisco because we’re not taking on the tradi- is partnerships. We’ve gotten closer to the If you ask 10 people what SDN means tional enterprise market. But at other levels big data companies like Cloudera and the they’ll give you 10 different answers. But I feel like I learned a lot about what not to virtualization companies like VMware. if I had to describe it in one or two words, do and how to do [things] better, by being Because we are best-of-breed, we are in I would say: open and programmable. more application-focused and really taking some ways less of a threat and more of a There’s been so much vendor lock-up in advantage of the market disruption from partner to many. The security companies, networking, with the huge operational cost enterprise to cloud, and then to big data like Palo Alto Networks, the application of being locked in with one vendor. That and network virtualization. Cisco, in my delivery companies like F5, view us as a mainframe model in networking is what view, will always be the enterprise market friendly face. SDN is challenging. Why is SDN fever and leader. Arista is inspired and aspires to be hype so high? People are sick and tired of the cloud networking leader and really be a What do you do with Cloudera? the vendor lock-in and proprietariness and complement to Cisco. We have actually installed joint networks they are looking for a movement. Now, how for big data together. The biggest issue and when will that movement happen? Other networking companies have with using these kinds of direct-attached Like anything, you have to be pragmatic tried to address the needs you Hadoop systems is that you have to have a about that. When the hype is so high the have described. Let’s take a case network with fast failover characteristics, market isn’t that big, right? But in my view, in point: Juniper seemed to be tar- the right buffering characteristics, and SDN has a tremendous opportunity to geting the same kinds of problems, you actually almost have to have a Hadoop succeed if we’re pragmatic about the use so why did the company struggle tracer-like function between the storage cases. Is SDN OpenFlow? No, it’s not just with its QFabric? and the network. Because, remember, all OpenFlow. Is it OpenStack? No, it’s not just Without making it specific to one ven- of a sudden you wiped out the concept of OpenStack. Is it VXLAN or network vir- dor, I would say three or four years ago, a storage-area network, but you still need tualization? Understanding what you can you rightly pointed out that the market to have the resilience of a storage-area actually do with SDN is the key here. was very crowded. There was Juniper, HP network. We’re in several joint customers looked like they were coming on strong, together, particularly in mission-critical Let’s talk about that programma- there was Force10 that then got acquired financials. bility aspect. When I talk to people by Dell, and then there was Brocade that in the market, there are some had acquired Foundry. They have taught I want to go back to EOS for a who are really excited about that me that focus is important. We stayed re- bit. There is sort of this myth of piece, but others who think there lentlessly focused on building a standards- the unified operating system in are only a limited number of things based open architecture for data centers. networking, but competitors and you would ever want to program I think many of these vendors, they get customers are all running mul- the network to do. What do you distracted by growth. It’s a difficult call, tiple versions and flavors of these envision people programming the in the public market especially. It’s easy to OSes. What makes this software network to do? do when you’re private and you don’t have different? Why would somebody I think they’re both right for different to explain things to people. But do you go believe that you have a single reasons. In a traditional network, no one is broad and wide or do you go a mile deep? unified network operating system looking to toss their IP out any time soon, My belief is I am participating in one of where people have struggled with so whatever you do you’ve got to make sure the largest total available markets in the that? your IP network is up and running. But history of networking. The 10G market is We’ve proven it. Even though we’ve say you’re going to do a green field [instal- going to go from a couple of billion to $15 been here five years and we’ve done all the lation]. This is how the OpenFlow SDN billion in 2016. I should not be distracting software releases, we still have one single movement got started at Stanford. [Profes- myself with other markets. I should relent- binary image. Nobody has to get a Ph.D. on sor] Nick McKeown was doing this project lessly focus on doing my one market very our software releases because we’ve kept called Clean Slate, where he was telling his well. I would say that is the failing of many one single unified code base. Secondly, students: “Imagine a world with no IP. How competitors who don’t have that kind of when people play with the software, they would we define it?” That’s all great for focus. realize that: “Oh, this can do Python script- vision and strategy, but we’ve got hun- ing, I can write agents to it.” So the power dred of thousands of customers operating How has the strategy changed that we’ve given them to enable things that networks, so you have to understand that since launch? What have you even we don’t enable is like no other. Now, no matter how inflexible and how much of learned in that time period? I could tell you some people love it. The a headache your current architecture may We built a point product then and our engineering community loves it. The CIO be, TCP/IP does work. Then you look at the strategy has changed to a portfolio. We community is afraid of it. So we also have a model that SDN is coming up with, which have gone from a top-of-rack (48 ports) lock-down mode where we can have all the I call the controller, overlay controller to 384 ports in a chassis. Secondly, our security and people can’t go in and simply model. That’s a controller for OpenStack, software has evolved, not just in millions start writing. OpenFlow, network virtualization, each of lines of code, but in terms of the agility At the same time, I would say we are a of them is a use case — it’s a specific case and innovative features. We’ve been put- little bit of a Sybil with our EOS. One side where you need programmability. I agree ting out releases practically every quarter of us looks just like a Red Hat or Linux, and that you shouldn’t go and mess with the big since then, to the point that some of our the server guys have fun with it. They can IP network. If it’s not broken, don’t fix it. customers say: “Hey, slow down a little! do TCP and dump and bash. It’s a Linux Understand what are the use cases We’re unable to absorb it as quickly as kernel, right? Then the other side, for a you’re trying to augment with program- you put it out.” One customer said to us: Cisco administrator, looks just like a Cisco mability. I can think of three or four use “We have all these data centers and we’re network. When we work with VMware, we cases. One of the ones that we found, racking up our servers, we’re connecting talk straight into vCenter and vSphere. So I especially with OpenFlow, but also in IP it to a VLAN and we’re enabling DHCP, think the flexibility that we have offered in networking, is data tap aggregation. When this whole process is taking us two hours.” being open, at the same time not destroying you’re running at the 10G speeds we are, With our zero-touch provisioning we were the paradigm between the virtual admin everybody is looking for traffic visibility able to cut that down to 20 minutes. You and the sys admin for the server, and the and understanding what’s going on in the think of this in one data center and how network admin has been unique. network. You can build an out-of-band

www.networkworld.com february 22, 2013 3 controller with OpenFlow, whether it’s Big I don’t know if there are that many SDN enon and revolution is happening in Switch Networks or open source control- use cases, but I love it.” Arista feels very networking. lers from Floodlight or NEC, and have an fortunate that we got an eight-year head OpenFlow agent on our switches, and have start. We spent four years building the But does it limit the speed at which a very simple SDN network that’s highly software, four years commercializing it, you can develop your product? programmable, and still works in hybrid and now we’re sitting in the middle of an No, actually, it’s better. We were most mode with your IP network. That could be SDN momentum and a switch momentum scared when we were only beholden to one case. There’s the Nicera case, which is that puts us on the natural cusp, if you will, one vendor. The way our software archi- also, in my view, a programmable use case, where we can be working with the old and tecture works, 90% of our engineering is but it’s for strictly network virtualization. still developing the new. all software. Then there’s a driver layer. You keep the IP running, but you need a The driver layer is where we custom- network virtualization platform that can Does Arista have its own ize to the silicon. Today we operate with program your virtual switches, whether controller? Intel drivers, with Falcon drivers, with they’re VMware switches or open virtual No. Broadcom drivers, we used to have a switches from an OpenStack environment. company called Dune that got acquired by Today, this is literally like two islands; Is that something you’re going to Broadcom, you know, Marvel. The more there’s the physical switch and the virtual have? drivers, the more we have the roadmap switch. With Arista working closely with Our control plane and data plane will be and the R&D of all the silicon vendors, VMware and Nicera, we can transcend open, and we will work with all the control- because remember there are always risks. the virtual-to-physical islands, where ler vendors that you deploy. HP Opsware, The biggest risk with the silicon vendors every VXLAN vSwitch port automatically vSphere VMware, Nicera, Big Switch, is if they miss a cycle. They don’t miss it maps to a hardware port. Now you’ve got Floodlight, NEC, IBM Tivoli, EMC Smarts. by one or two weeks. They often miss by network virtualization not as two sepa- These are all to us controllers. We don’t one or two years, whether they work for rate failure domains, but transcending view ourselves as the management expert. you inside the company or outside. We virtual-to-physical, potentially even to a Now, if you want to go develop specific use like the fact that we have drivers and have cloud, architecture down the road. This is cases for a controller on our switch, we developed drivers over time for all these in a VMware environment, but there’s no can do that too. And people do that. Like chip vendors. reason you couldn’t do it in an OpenStack for example, we worked with Splunk very environment with quantum plug-ins as closely to develop a configuration manage- What’s ahead for Arista in 2013? well. So that’s another use case. ment tool with them. Expect to see more capacity, more software, more of everything, and the re- Do you believe SDN fundamentally If you’re a CIO or a senior IT execu- placement of our older products. Second, changes the competitive land- tive, and you are on the sidelines expect us to continue to be application scape? looking at SDN and wondering what focused, because that’s where we think we Yes, but I believe the controller vendors it’s all about, what’s the 30-second get appreciated. If it’s a straight, tradi- in isolation won’t succeed. The networking answer on why they should care tional network, where nobody gets fired vendors, if they get defensive with just IP about it and start thinking about for buying IBM or Cisco, that’s not where won’t succeed. You have to have the Arista it now? Arista shines. Arista will continue to pro- view, which is have two personalities. Because it unshackles their traditional vide thought leadership, like we did with Work in a controller mode, but work also networking decisions from their appli- high-frequency trading, and like we do with your IP network. That’s the mistake cation and helps them develop new use now with big data and network virtual- I think SDN is making, in that they’re cases that they couldn’t do before, or only ization. We’re very excited when we think thinking of it as only green field. In reality had to do with traditional vendors. The of the number of things we have to do in there’s a whole world of IP there that you key for them is to understand what prob- the software and application world, and need to work with while you’re trying to lem they are trying to solve. Solve that prioritizing them, there’s so much ahead develop the use cases. out-of-band and work with the existing of us. [network]. Server virtualization was essen- Last question: the classic elevator tially owned by VMware and the Using merchant silicon versus situation. You’ve got 15 floors with software vendors, not the server custom ASICs, is there a limit to a CIO. You tell them what about vendors. Who ends up owning SDN? where that can take you? Arista? [The server vendors] got defensive. If the I think custom ASICs have fallen behind We’re the best-of-breed data center and networking vendors do not embrace it be- by several years, and unless a vendor cloud networking alternative in the market. cause it requires a new software paradigm can build and compete against merchant We’re not a cost center, we are an enabler that they haven’t built, then they stand to silicon, there’s no point in doing custom of applications. There are people in high- lose, because over time there will be more ASICs. It’s something that the established frequency trading who have told us: “We and more use cases that disrupt them. vendors have to come to terms with, just make $2 million a day on your boxes.” In Customers like that. I was talking to a cus- like the server community did. There was a nutshell, that’s what I would say. We are tomer today who said: “The most liberating a time when SPARC was sparkling for the thought leader in cloud networking, we feeling for me now is I can be multi-vendor, Sun, and then everybody had to embrace enable the applications and we save you and it hasn’t cost me that much money. the x86 Intel world. A similar phenom- CAPEX and OPEX.

R&D and Headquarters 5470 Great America Parkway • Santa Clara, CA 95054 • Tel: 408-547-5500

Email: [email protected] US & North America Sales: [email protected] Latin America Sales: [email protected] Europe, Middle East & Africa Sales: [email protected] Asia-Pacific Sales: [email protected] Japan Sales: [email protected]

© Copyright 2013 by Network World, Inc., Southborough, MA 01772-9108 • Posted from Network World.com • Trademark is owned by International Data Group, Inc. #C18171 Managed by The YGS Group, 800.290.5460. For more information visit www.theYGSgroup.com/content.