High Speed Networks

On to Petabit Networks

High-speed networking authority Craig Partridge says gigabit rates aren't nearly high enough

ISPs are reporting some big numbers when it comes to the traffic their networks are handling. Growth rates are anywhere between 15 percent and 30 percent per month; stretch that over 12 months and annual growth rates are 500 percent to 1,800 percent.

So maybe these rates won't last forever. But improvements in such access technologies as xDSL (digital subscriber line) suggest we're in the midst of a continuing trend. That means carriers will have to build bigger trunks. How big? Try petabit-per-second. How soon? By 2007, if growth proceeds at just 15 percent per month. And if it continues at 30 percent, they'll be needed by 2002.

To put the capacity numbers into perspective, consider that a gigabit is equivalent to 109 bits and a terabit to 1012 bits. A petabit, on the other hand, is 1015 bits of capacity. And those on the leading edges of research already are throwing around some much larger figures, like exabit (1018 bits), zettabit (1021), and yottabit (1024).

Which begs the question: How are we going to achieve these kinds of rates? The simple answer: No one knows. But by doing a little bit of extrapolating, it's possible to get a clear read on the approaches likely to pay off. And along the way, net managers get an enlightening tour of gigabit and terabit networking.

As Goes Fiber...

Progress in networking is usually driven by progress in transmission speeds. And when it comes to high-speed networking, this observation can be summarized as follows: "Where fiber goes, everyone else follows." So let's look at the state of fiber.

The past couple of years have seen two major trends. The first is the ongoing improvement of WDM (wavelength-division multiplexing) technology. Nippon Telegraph and Telephone Corp. (NTT, Tokyo), which boasts the leading WDM research lab, is using the technology to pack larger and larger amounts of bandwidth into fiber. The company has moved way beyond the terabit level, and fiber itself has a theoretical capacity of around 50 to 75 Tbit/s.

In the commercial domain, WDM is so cost-effective that standards work on faster serial speeds in fiber (like OC192c [10 Gbit/s] and OC768c [40 Gbit/s] ) has pretty much come to a halt. Some carriers have concluded that using WDM to send multiple OC48 (2.5 Gbit/s) channels is more economical; in fact, MCI Communications Corp. (Washington, D.C.) already uses an OC768 link over WDM.

Right now, the commercial standard is 8 or 16 channels per fiber, a density referred to as Dense WDM (DWDM). But researchers already are packing thousands of channels into one fiber, which means DWDM will soon be supplanted by even denser technologies—and marketing departments will have to come up with other terms to describe them.

The second fiber trend is continued improvement in analog signaling techniques. As we all know from modem technology, it's possible to send more than 1 bit per pulse by varying the amplitude in the pulse. In a fiber optic transmission, the norm was just 1 bit per pulse—until recently. Now researchers are sending as many as 3 bits per pulse, effectively tripling the theoretical capacity of a fiber to about 150 Tbit/s.

The innovations of the late '80s and early '90s have moved quickly into products. The most important of these are probably the improvements to the optical amplifier. Signals in fiber degrade over distance and need to be amplified every 100 kilometers or so. This used to be handled by optical-electrical amplifiers, but these devices put major limitations on how many channels could be signaled through a fiber and how fast it could be done. Upgrading a link meant upgrading all the amplifiers in the fiber path.

But now, with optical amplifiers, that's not a problem. Transmitting more channels simply means upgrading the equipment at the ends of the fiber path—and the optical-electrical amplifiers already in place can stay in place.

All of this bodes well for petabit networks. Indeed, that small fiber bundle a company might already have has (in theory) a petabit of capacity. Customers will just have to wait for "ultra dense" WDM (or whatever it will be known as) to use it.

Fast Packets

Switching technology is booming. Commercial routers and switches now incorporate switched backplanes capable of moving 20, 40, or even 100 Gbit/s. And researchers are doing even better: At Stanford University, Nick McKeown—with help from Texas Instruments Inc. (TI, Dallas) and Cisco Systems Inc. (San Jose, Calif.)—is putting the finishing touches on the Tinytera. It's a near-Tbit/s switch the size of a soda can, made using CMOS technology. Optical packet switching isn't capable of those speeds just yet, but it's improving quickly—and it will be available if and when electronic switching runs out of steam.

Meanwhile, routing technology isn't the laggard it once was. With Internet data rates increasing, established vendors, startups, and research labs have made 1997 a year of technical innovation. Commercial routers can now forward well over 10 million packets per second (pps). (As a rule of thumb, for every gigabit of bandwidth, a forwarding capacity of 0.5 million to 1 million pps is needed, assuming average packet sizes in the 128- to 256-byte range.)

Even more stunning, the least expensive routers can forward packets almost as fast as the most expensive routers, thanks to ASIC technology (although the most expensive routers can do more per packet, such as applying firewall filtering). Industry and university experts generally agree there are only a few remaining obstacles to building terabit routers.

But a terabit is a factor of 1,000 short of a petabit. How do we get to petabit speeds? Here's where Moore's Law can help out. It predicts that by 2002 chips will be about 10 times faster. Smart engineering can often boost that by a factor of two or better. That means we only have to go 50 times faster. Recent research on routing table management at Washington University (St. Louis) and Lulea University (Lulea, Sweden) is promising a factor-of-10 performance improvement, which means we only have to go five times faster. And parallelism can probably buy us that last factor of five.

All of that makes it seem as though a petabit network isn't outside the realm of possibility. The question is, where is all the data needed to fill that network going to come from?

Let's look at access technologies. We all know that wireline technology is improving. Cable modems and xDSL both furnish megabits to the home. It's worth remembering that 100 million homes and offices wired at 10 Mbit/s is a petabit of edge capacity. Also keep in mind the work of UCLA's Leonard Kleinrock, whose latest statistics show that 90 percent of all LAN traffic has a destination off the LAN (a stunning reversal of the trend just a couple of years ago). So much of that edge capacity shows up in the middle.

Wireless is improving, too. A recent Ph.D. thesis at MIT showed that by making better use of power management in radios, it's possible to deliver 200 Mbit/s per user. Also worth keeping in mind is that we'll soon be able to attach little wireless antennas directly to silicon chips, making all sorts of new communications patterns possible—not to mention very cheap.

And then there's always satellite technology. Experimental satellites now have channel speeds of up to 622 Mbit/s.

In short, it probably isn't too early to start thinking about what comes after petabits. Personally, I get stuck around yottabits. A fiber bundle capable of carrying a yottabit per second is about 12.5 meters in diameter—the size of a subway tunnel. Routers would need to buffer more bits than there are atoms in the largest chip. But that's OK: Even if growth continues to boom, we won't get to yottabits for about 15 years.

[ Home ]

[ Registration | Subscriptions ]
[ Contact Us | E-Mail ]