There has been a fair amount of industry discussion about the need to delayer enterprise networks to simplify them. What's your take?
Yen: A three-layer structure is conceptually very helpful, particularly in the data center. You have access, aggregation and the core. But with today's technology, particularly what Juniper is offering, we believe pretty much all data centers -- except a handful of large core operators such as Amazon and Google -- can be supported using a two-tier structure rather than resorting to three tiers.
When you reduce a tier, a layer, by definition you save in cost, you save in power and you have better latency that translates to better performance. So this is why we're winning a number of data center deals -- including the New York Stock Exchange -- with our two-tier approach, the fixed configuration EX3200 and EX4200 in the front, and the modular EX 8200 switches in the back. And very frequently, MX will become part of the equation when the customer looks at edge routing.So you squeeze out the aggregation layer.
Perdikou: You always have the access layer because that's the highest volume, but instead of using a middle layer to reduce the number of ports, if you have sufficient port counts at the core, which is ultimately the fabric providing the switching function, you can subsume the aggregation requirement.
Of course, in order to achieve that, if the total number of ports or lines after the access layer is still huge, you need a bigger structure. But the reason we can always accomplish this in two tiers is, at the access layer we offer software technology we call virtual chassis. This is a way to take multiple access switches -- up to 10 -- and aggregate the collection as if they were one logical switch. That simplifies the management but, more importantly, once they aggregate this way you reduce the number of uplinks.
So, it reduces the number of uplinks and reduces the pressure to have a more complicated higher level structure. Then you bring in the high port count modular switch -- which can serve the combined role of core and aggregation.How big a company can this architecture support?
Yen: That type of structure can cover up to 4,000 or 5,000 servers, pretty much all major businesses.
Perdikou: If you think about the ones that it doesn't fit -- like the online business, what we think of as content service providers -- the two-tier approach with very few exceptions can go from the smallest to the largest.
OK, that's two tier. With the Stratus effort you briefly outlined last year you're talking about moving to a single layer, right? Give us the thumbnail.
Yen: We haven't told the world too much, but last year we publicly disclosed we are undertaking a project called Stratus. Marketing picked the name because stratus is a single layer of flat cloud, and it implies what we are trying to achieve. You can think about it as highly scalable, from supporting several hundred 10Gig E ports to tens of thousands of 10Gig E ports, and all at line rate, so it's not a heavily oversubscribed type of implementation.
It's one architecture, very scalable, any-to-any (any ports can connect to any other ports), it's homogenous, fair (there is no bias of, if you go this way or that way it's faster), and most importantly it is lossless. In the Internet you're allowed to drop a packet upon heavy congestion. But in the data center, for communication between servers or between servers and storage, it's totally unacceptable.
Stratus is also a converged fabric, which means instead of the current practice of using Ethernet for IP traffic, Fibre Channel for storage traffic and Infiniband as a low-latency technology, Stratus will support a converged Ethernet fabric that will support all kinds of data center traffic. And on top of that, Stratus will guarantee a very low, worse-case latency and be competitive to InfiniBand.
If your data center is small, you have 50, 200 servers, then the intranet is nothing but the interconnect among servers and storage. The total cost is low, power consumption is low, latency is relatively low. You don't really care.
But then as your scale grows so does the accumulated latency. And it's only going to get worse with multicore, multithreaded microprocessors that can significantly boost the total throughput a server can drive. And with virtualization you want to run four or eight or 16 virtual machines on one physical machine. Suddenly each server is faithfully driving its gigabyte Ethernet lines -- or 10Gig lines -- near the line rate most of the time and the cost of all of this makes the network a first-class citizen in the data center, just as much a concern as the servers, just as much a concern as the storage.
So, that's what inspired Juniper's Stratus vision. It started in Pradeep Sindhu's organization [Juniper's CTO and founder] and Pradeep has been nurturing and fostering the effort and now it has grown into a full-scale project. We've been working on this in total probably more than two years.
How will it be productized? Is it an answer to Cisco's Nexus stuff?
Perdikou: It's a natural thing to try to compare it because Nexus is Cisco's latest, newest push. There are a few areas we are thinking in common. For example, Nexus is striving toward converged Ethernet within the data center, and in that we are completely aligned. We believe the data center is heading in the direction with unified, converged traffic. If you build a 10-lane highway from downtown to the airport, you want all ten lanes usable by different kinds of vehicles -- whether they are four-door sedans or a pickup truck -- so you can smoothly, flexibly handle the traffic volume.
But Cisco's approach is very incremental. You look at the highly touted Nexus 7000, and pretty much all the sales right now are to replace aging Catalyst 6500s. There is very little real converged traffic being driven on the Nexus line, even though it has been in the market for a couple of years. And you look at the product implementation, it's a very incremental evolution from today's switch. It's nowhere near the fundamental architecture change as in Stratus.
Yen: What Stratus presents to the customer is intended to be a very simple image: a very scalable single data center fabric. Under the hood, Juniper is throwing all the experience we've gained over the last 13 [or] 14 years of cutting-edge switching technology and Internet core routing.
Such an offering takes a lot of effort -- the silicon investment, the hardware investment, multiple layers of software. Now, come back to Juniper and Cisco. If you look at it from the technology capability, Cisco has a lot of talented engineers, so we believe our respectable competitor will have the capability. And if you look at it from the resources perspective -- the ability to invest and create such a product -- they are resourceful enough. But from the business perspective, there's a significant difference between us.
Cisco is already very dominant as a data center intranetworking provider, while Juniper has very little share. So where is Cisco's motivation to say, "Let's introduce a fundamentally new architecture to revamp the whole thing." Moving the money from your left pocket into your right pocket is not going to make you any richer. So, the business incentive for Cisco to significantly invest in a very fast pace of evolution is not there.
Ever since we publicly disclosed Stratus last February we are getting overwhelming interest from all over the world. Everybody wants to call and talk to us. It's similar to -- before I joined Juniper I spent 20 years at Sun Microsystems -- when Sun invented Java in 1995. The phone was ringing off the hook because everyone wanted to know something about Java. And right now, while it will still take some time for Stratus to become available, we are actually enjoying such publicity.
What is the timeframe for Stratus?
Yen: It will take a couple of years. But whenever certain pieces of Stratus technology become available we will offer it on top of today's product.
Stratus inside? (Laughing from group).
Yen: We may not give it the Stratus label but we will apply some pieces of technology when it becomes available.
Is Stratus similar to what Jayshree Ullal and Andy Bechtolsheim are up to at Arista?
Yen: Certainly they have gotten a lot of publicity and they have done a respectable job considering their size. But if you read some of their speeches they say they're pretty much providing the access layer. And when asked how about the core, they actually say the customer will have to go through Cisco or Juniper to complete a whole configuration. They know their limitations.
Shifting gears to computing, what do you make of Cisco's Unified Computing System?
Yen: Having come from the server world, UCS is just the next generation of blade servers. The vendors that reacted to the move were the server companies, IBM, HP, Dell. Cisco literally invaded their market. But instead of joining the very low margin X86 vendors and building conventional X86 servers, they carved out a portion of the network interface, a portion of the storage interface, and then adopted the latest Intel multicore, multithread microprocessor and, with their relationship with Intel, came up with an innovative approach to provide more memory on their blade, which is conducive to the number of virtual machines they can support. Then they laid in the VMware virtualization software and the BMC management software.It's a local scale system integration job. But by doing that integration, obviously, they eliminated some unnecessary hardware/software which reduces the total cost of such an aggregation. And with the reduction of that cost, they could give the customer half of the savings as an incentive and keep half of the saving to raise the margin.
So, it's an interesting move on their part. Unfortunately, they enter territory they don't really know that much about. Getting into servers is not just putting the hardware together, but most seriously, they are hurting their partnerships with those system companies.