The internal bottleneck: Terabit switch routers are coming of age

What is the recipe for the fastest network possible? Most service providers couple dense wavelength division multiplexing with high-speed fiber links. But pumping up the link speed and capacity is not enough. If you don't install the right switch, you'll still have a bottleneck.

That's the problem stifling today's networks, said Mukesh Chatter, president and CEO of Nexabit Networks. "Bandwidth access is increasing at the core, and the core has to keep pace with it," he said. "Congestion delays are a serious problem."

All new builds and upgrades combine fiber, WDM and routers. Each of those is increasing in capacity, but not at the same ratio, Chatter said. By mid-1999, fiber will be used 100 times more than it was three years ago. Some DWDM systems claim to support 128 channels on a single fiber. The advancements in fiberand DWDM provide 10,000 times more raw bandwidth than was available a few years ago, he said.

"But the third leg of the tripod, router performance, has only gone up by a factor of 10 in the last few years," Chatter said. On a new infrastructure build, he said, "not even a hundredth of 1% [of the available capacity] can be used because the router is the bottleneck."

Service providers are clamoring for faster routers, according to Chatter. That has led to the birth of several terabit switch router companies such as Avici Systems, Juniper Networks and NetCore Systems. But the problem is that those companies don't have product available yet, he said, and what they market isn't a true terabit switching fabric.

Of course, Nexabit isn't shipping product yet, either. Its NX64000 multiterabit core switch router will go to beta testing this year. Juniper is currently beta testing product, and others vendors are still in development.

But Chatter takes offense when Nexabit is lumped in with other terabit switch router companies. He claims that the NX64000 performs at true terabit levels-1.6 Tb/s per slot.

"It's like if you have a 300-baud modem and you compare that with a cable modem," Chatter said. "We define a 1000 Gb switching fabric as the minimum requirement to be considered a terabit router. Juniper is 40 Gb; that's 4% of a terabit router. Cisco is 6%. Avici's got 10% of a terabit router." Nexabit makes a multiterabit router, he continued. "That doesn't mean we don't support less. We can support 80 Gb today, and the cost doesn't go up when you go up to a terabit."

Competing vendors don't agree with Chatter's suggestion that the switching fabric has to start at a terabit. Customers are more interested in I/O and port density, said Pete Chadwick, director of marketing for Avici.

"The idea that a router is a single chassis is ridiculous," he said. "Carriers are used to building out services with multiple bays. The issue is how you're going to scale to handle the growth that our customers are currently seeing."

Juniper Director of Marketing Joe Furgerson acknowledged that its device uses a 40 Gb switching fabric, but the fabric alone does not make the router. "The key issue is software compatibility," Furgerson said. "Cisco has 90% market share in this space. It's not because they have the best hardware, but they set the pace for software innovation."

Administration of the switch, as well as how it connects to other switches within the core and at the access layer, are crucial to the Internet service providers that Juniper targets, Furgerson added. Juniper supports a wide range of interfaces to the core and access layers and will boost its switching fabric in the future as needed.

Some lower-speed devices use clustering to improve performance, pointed out Dave Passmore, president of NetReference. "The problem with clustering is that a lot of traffic has to go through multiple hops within the cluster. Each router is wasting a lot of capacity just to talk to other routers in the cluster. You'd really like a single switching fabric that provides high performance, and that's where Nexabit seems to be."

But performance alone doesn't make the device. Timing is important, too. Juniper was late to market, which could hurt its ability to eat into Cisco's market share, Passmore said. "Juniper is not that much faster than a Cisco 12000, so if you're already a Cisco customer, you have to ask, 'Do I really want to give up Cisco and take a risk with [a start-up] like Juniper?'"

But if the switch speed is 100 times faster, as Nexabit claims, "it might be worth it," he added.

The adoption of terabit switch routers will make bandwidth a price-driven commodity, Passmore said.

"If you already paid for the [WDM] and the fiber, all you have to add is the high-speed router," he said. "Now you can drive your competition out of business by being much more efficient than they are. You can lower your cost of bandwidth way below that of the competition; obviously this could start a price war."

The caveat, of course, is that terabit switches only address backups in the core network.

"We haven't solved the last-mile bottleneck," Passmore said. "As a result of WDM and high-speed routers, we could see a glut of bandwidth in the wide area, but end user customers may not see the benefit of this because of the last mile and local distribution."

What about skeptics who say terabit routers are not necessary in today's networks? "Those who don't have it say there is not a need for it," Chatter said. The math proves it. If bandwidth is growing 10 times each year, and today networks require 100 Gb/s, "What is required in a year?" he asked.

ATM IS LOOKING GOOD According to research firm Vertical Systems Group, sales of ATM equipment to carriers are expected to grow at a compound annual rate of 37% between 1997 and 2001. ATM sales to ISPs will grow more rapidly, at a rate of 54% annually during the same period. This year, the firm projects that ATM equipment revenues will surpass frame relay equipment revenues.

A CLUSTER OF QOS HydraWeb has announced the HydraHydra HH100, which combines application-aware routing with cookie-based prioritizing capabilities and a Java-based global service management tool to provide site-level load balancing across a clustered environment.