Switching Silicon Goes Scaleable

The big problem with routers and switches today is that their lifecycle is too short. Every two or three years, the equipment becomes outdated, and since it cannot be upgraded further, it has to be stripped out and replaced. That's costly, both for service providers and for systems vendors, which have to develop new products on a three-year cycle.

So says Eyal Dagan, a routing expert who previously founded Charlotte’s Web Networks Ltd. and who now heads up Dune Networks, a switch fabric startup that claims it can address this problem (see Dune Drops Stealth Shield).

Data equipment goes out of date very fast, he says: As port speeds increase, more ports are needed to address more customers, and new service schemes such as virtual private networks (VPNs) are introduced. Voice equipment, in contrast, survives a lot longer because the basic nature of telephone calls stays the same.

The upshot is that the moment a system vendor completes a router or switch product, they have to start a new one from scratch. Given that it takes an investment in the region of $100 million to develop a totally new system, it's very difficult, Dagan contends, for system vendors to generate enough revenue in just two or three years to recoup that investment.

In his view, the root of the problem lies with the switch fabric. Some equipment upgrades are possible by adding new line cards, but ultimately the system is limited by the capabilities of the switch fabric.

"The switch fabric is the core thing, whether its designed by you or is an off-the-shelf fabric from a vendor," says Dagan. "You are bound by the limitations of the switch fabric -- its port count, port rate, and service scheme."

Dagan founded Dune back in 2000 when he realized that all the new generations of switch fabrics from chip vendors were being built on the same principles as previous generations, merely offering fixed improvements in port count, line rate, or service scheme. His idea was to develop something more flexible, which could be scaled to port counts, port rates, and service schemes that have yet to be invented.

The result -- a switch chipset called SAND -- is currently in the labs being tested and should start sampling before the end of the year.

It sounds like a cool idea, but at this stage Dune is only giving out partial details about how it will make it happen. However, analysts who've been treated to a closer look at Dune's product appear to be buying into the story.

"Dune Networks brings the most scaleable architecture to the fabric market" says Jag Bolaria, senior analyst with the Linley Group. "SAND allows OEMs to extend their product life cycles from the legacy three years per data-communication product to a life cycle of seven years and more. This significantly cuts costs involved in rebuilding platforms, de-bugging, and retraining for both the vendor and the service provider."

Like many switch fabics, the chipset comes in two parts: a scheduler, which sits on the line card, taking traffic from a network processor and sorting it into queues; and a switching element that sits on the switch card. Dune claims the chipset will be highly integrated, requiring only 1.125 chips per 10-Gbit/s port. Later next year the startup plans to introduce a new version of the fabric with three times the density, or 0.375 chips per 10-Gbit/s port.

Dune claims that its chipset differs from most others out there in that it has a distributed architecture. That means carriers can start with one chassis, and when it reaches full capacity, they can link it to another chassis in such a way that the switch still behaves in a non-blocking way. "Fully distributing the intelligence to the line card is part of the secret," says Dagan.

Of course, the idea of making distributing routers is not new among system vendors. Juniper Networks Inc. (Nasdaq: JNPR) announced such a product earlier this year (see Juniper Goes Terabit With the T640). Cisco Systems Inc. (Nasdaq: CSCO) is also rumored to have made several attempts to develop a distributed architecture "huge fast router." The internal development teams of these companies represent serious competition to a startup like Dune.

It's also worth pointing out that the core router market is in poor shape right now (see Optical Oracle: Core Incompetence? ). Many core router startups have not stayed the course, including Pluris, which was developing a distributed architecture router (see Pluris Shutdown Confirmed).

Dune's Dagan counters by saying that his company's chip is applicable to all parts of the network, not just core routers. It's a common misconception that only big boxes can benefit, he says. "Saying it's distributed doesn't mean you have to build a big device. You could start with a pizza box."

He also points to the fact that a carrier -- SBC Communications Inc. (NYSE: SBC) -- has invested in the startup as proof that the product addresses a real need. Dune raised $24 million in its first round (see Dune Digs Up $24M ).

Of course, even a good story and a good product may not be enough to guarantee success. At last count there were at least 20 other switch fabric startups out there, and, says Jag Bolaria, they all want a piece of the same pie.

— Pauline Rigby, Senior Editor, Light Reading
MrLight 12/4/2012 | 9:35:57 PM
re: Switching Silicon Goes Scaleable Since DUNE Networks is a member of the Network processing forum http://www.npforum.org/about/m... I am assuming they know what NPUs are out their and they feel they can do a better job.

"The SAND chipset comes in two parts: a scheduler, which sits on the line card, taking traffic from a network processor and sorting it into queues; and a switching element that sits on the switch card." Is a classic NPU architecture.

An interesting article to read is "Advantages of buying off-the-shelf NPUs" Feb.2002 by BY PAUL DeBEASI, ONEX Communications,a TranSwitch company at http://icd.pennnet.com/Article... .

Basically -
1.Product time-to-market
2.Reduced project risk
3.Simpler project scheduling
4.Reduced dependency of embedded software verification on hardware verification
5.Reduced product-testing and verification phases 6.A big, complex, lengthy design program could put the overall platform [the actual product]program at risk.
and I like this one...
7.By purchasing best-in-class merchant silicon, systems vendors can better focus on their core strengths and deliver a product with unique performance enhancements that makes it superior to their real competitors-other equipment vendors (not chip makers)
and this one ...
8.By leveraging merchant silicon, employees can focus on the projects and technologies central to the company's core competency and central to the platform project.


However the article is only partially correct since if the SAND chipset is DUNE NETWORK'S secret sauce, as Cisco's internal NPUs are theirs, than DUNE NETWORKS is right in rolling their one. Time will tell.

MrLight :) shedding light on the subject
Tomahawk 12/4/2012 | 9:35:54 PM
re: Switching Silicon Goes Scaleable Actually, the answer is "No."
fabric_man 12/4/2012 | 9:35:38 PM
re: Switching Silicon Goes Scaleable Taken literally, this article has disturbing implications. If all, or even most, of the scheduling intelligence is on the line cards, then all of those line cards have to communicate with each other in order to reach a sensible scheduling decision.

This scheduling communication must necessarily flow through the fabric itself. If it shares the datapath with real traffic, then customers are not getting the true BW they think they are paying for. If it flows out-of-band vs. the datapath, then that's better from BW-preserving point of view, but there have to be a lot of links between each line card just to carry this scheduling information. Those links aren't free.

Another issue is cost. It's all well and good to talk about low chip counts for 10G of switching. 0.375 chips per 10G would in fact be very good density.

But, the key is, what's the chip count for 1G of switching? There isn't much call for 30G or 40G fabric port devices unless they can switch (and apply flow control) on very fine grained lower-BW traffic. Most of the traffic out there is GbE or even slower. What if there is only 4G worth of traffic to switch on a line card? Why do I need to pay for a 40G fabric port?

The question to ask is, what's the cost per Gb of switching (not 10G) and what's the cost if there are only 2-4 Gb to switch? If the answer to this question is good, then the product has a chance of actually selling something.

The founders' prior association with a terabit router company may have given them the impression that bigger is always better. Low-cost is better. Bigger is just bigger.

-- fabric_man
Sign In