The throngs invading Las Vegas this week for NetWorld+Interop 98 should have no trouble finding vendors hawking Layer 3 switches; in fact, getting away from them may be the problem. ATM had its turn, as did 10M and 100M Ethernet. Now it's routing switches that command the spotlight.
With this development, we've in a sense come full circle - from routing to bridging to switching and back to routing again, given that Layer 3 switching is basically routing repackaged. It seems appropriate, then, to call a timeout and assess this switching phenomenon - to examine how far switching has brought us, where we stand today, where the technology is poised to go and how we're going to get there.
Today's networks are a far cry from what experts envisioned in 1990, when neat and tidy visions of end-to-end ATM connections and voice/data convergence were dancing in their heads. Ethernet switching was just a stopgap that would extend the life of aging technologies such as TCP/IP long enough for the new order to establish itself.
But the visionaries underestimated the inertia of the installed base and failed to learn the lessons of Open Systems Interconnection (OSI). Paradigm shifts bubble up from the bottom like natural ground springs; they aren't imposed from the top. As ATM slowly worked its way through standards committees, the Ethernet cadre rose to the challenge and in-creased network speed by two orders of magnitude in just seven years. Users started deploying ATM on the backbone, but the availability of a cheap, simple, high-speed Ethernet alternative effectively kept them from extending it to the desktop. And Ethernet will continue to dominate the desktop unless ATM prices come down and demand for ATM-enabled services goes up.
But Ethernet couldn't provide an end-to-end solution, either. While switches enabled the creation of larger Ethernet networks, they couldn't scale across the backbone because of management and spanning tree problems. It was a standoff until Layer 3 switching entered the picture.
"The advent of the Layer 3 switch is what made frame networks at the core possible," says Basil Alwan, vice president of product management at Bay Networks, Inc., in Mountain View, Calif. "You get one of the main benefits of ATM - you can switch or route equally fast. That has important implications for how you can build networks."
What's in a name?
Switching is one of the euphemisms that abounds in the politically correct 1990s as people use words to evade reality. Layer 2 switches are really multiport bridges, but routing was in, and the "B" word was out when the EtherSwitch was christened in 1990 (see story, this page). Similarly, the "R" word had become synonymous with slow, complex and expensive by the mid-1990s when the first Layer 3 switches - really hardware-based routers - appeared.When bridging was invented in 1983, it outperformed and displaced the single-protocol routers then being used. But bridges propagated broadcast storms, so they were replaced by multiprotocol routers or put into silicon and called Layer 2 switches. Now routers are out of vogue, so they are being implemented in hardware and referred to as Layer 3 switches.
"Routing is not bad, it just needs to be a lot faster and cheaper," says Drusie Demopoulos, vice president of marketing for Foundry Networks, Inc., in Sunnyvale, Calif. Although Layer 3 switches generally support only IP, Layer 3 routing switches are an order of magnitude faster than the traditional software-based multiprotocol routers that put Cisco Systems, Inc. on the map. And while a routed 100M bit/sec Ethernet port is priced at about $750, a port on a traditional router can run $5,000 to $7,000.
The price/performance gains from replacing traditional routers with Layer 3 switches are thus much greater than those obtained by moving from hubs to Layer 2 switches. Industry observers think the price difference between a switched and routed port will drop to zero within two years. Extreme Networks, Inc., of Cupertino, Calif., has already announced pricing of $100 per port for switched and $150 for routed 100M bit/sec Ethernet.
Currently, vendors are each using their own custom Application Specific Integrated Circuits to manufacture Layer 3 switches. "The next trend is putting Layer 3 switching into standard silicon," says Diane Myers, a senior analyst at In-Stat, a consultancy in Scottsdale, Ariz. "We should see some announcements from the semiconductor companies within 12 months." This commoditization will help to reduce the cost of building Layer 3 switches (NW, March 2, page 39).
Meanwhile, Layer 3 switches can reduce the cost of network ownership by making it a lot easier for administrators to do moves, adds and changes. They don't need to know what physical port users are at.
Consequently, "[virtual LANs] are much easier to implement in Layer 3 switches," says Ori Bendori, chief technology officer for LANNET, a subsidiary of Madge Networks NV, based in Tel Aviv. "They are almost transparent."
When creating and managing VLANs at Layer 2, administrators have to track the physical ports and media access control addresses of individual users. At Layer 3, VLANs can be implemented by assigning IP subnets to them.
Prospects for ATM
Given their obvious benefits, nobody really doubts that Layer 3 switches will do well. ATM is the big question mark.Until the recent emergence of gigabit Layer 3 switches, ATM was the only choice for upgrading core networks. And big risk-averse IT shops still see it as the safest bet - particularly if they are thinking about quality of service (QoS) and voice/data integration. QoS is embedded in ATM technology, and ATM is the telephone companies' preferred WAN solution.
In contrast, "Gigabit Ethernet hasn't really ramped up yet and won't have a solid QoS solution for several years," says John Armstrong, an industry analyst at Dataquest, Inc. in San Jose, Calif. "ATM is a much more mature technology."
But the best technology doesn't always win. Outside of FORE Systems, Inc., it isn't easy to find people who are bullish on ATM. Some say ATM has achieved a critical mass of adoption that will carry it forward, albeit with its growth checked by Gigabit Ethernet. Others are less optimistic.
"Despite record sales, ATM on the campus is doomed," asserts David Passmore, president of NetReference, Inc., a consultancy in Sterling, Va. LAN bandwidth is virtually free once the equipment has been installed because there are no ongoing line charges. Consequently, it makes more sense to over-provision the network than to mess around with QoS parameters.
"ATM had an opportunity and blew it," agrees Brendan Hannigan, an industry analyst with Forrester Research, Inc. in Cambridge, Mass. "It's too complex and too expensive."
Even committed ATM users are feeling beat up by all this negativity. When Chesapeake Energy Corp. in Oklahoma City decided to upgrade its core network with ATM last spring, network supervisor Bryan Sagebiel went to Interop in
Las Vegas, eager to attend tutorials and learn more about the technology. But instructors, far from sharing his enthusiasm, were "kind of dogging ATM."
"I don't understand why," Sagebiel says. "I was skeptical at first. I figured, Ethernet is fine, so why spend the money? But since the migration, I think ATM's the greatest thing since sliced bread."
Infonetics Research, Inc. in San Jose, Calif., recently completed a study of users with high-speed LANs that suggests Sagebiel has plenty of company. The number of respondents using ATM doubled over a similar study conducted the previous year.
The respondents plan an average of 24 ATM backbone ports by 2000, amounting to 22% of their total. "So a lot more people are using ATM," says Mike McConnell, director of LAN programs for Infonetics.
Wrong assumptions
Industry pundits say a lot of ATM's problems stem from the fact that its original objectives are no longer valid. Bandwidth was assumed to be very expensive, so the idea was to micromanage it to squeeze out every drop. In a time of cheap bandwidth, this results in a lot of unnecessary effort and complexity.ATM was based on the premise that it would go end to end, but it failed to get all the way to the desktop. Consequently, a lot of its services have little utility. Also, the mixed ATM/Ethernet environment results in segmentation and reassembly overhead that can't be tolerated in environments supporting compute-intensive applications such as imaging and simulation.
There also was a notion that fixed cell sizes were needed for switching to be efficient and cost-effective, but silicon advances changed the equation. "And frame switches are proving to be faster than cell switches, so ATM might yet lose out in the WAN," NetReference's Passmore says.
But a lot of the blame for ATM's bad reputation was earned in the early days, when users were struggling to implement first-generation technology.
"The older equipment used [permanent virtual circuits] instead of [switched virtual circuits]," says Richard Sweatt, director of marketing for Hitachi Internetworking, a division of Hitachi Computer Products America, Inc. in Santa Clara, Calif. Network managers had to run around manually configuring every desktop and endpoint. PVCs worked with Private Network-to-Network Interface's (PNNI) predecessor, the Interim-Interswitch Signaling Protocol. Administrators had to configure by hand each switch's address and the hierarchy of the switch interconnections - local address, group address and path to the backbone.
This didn't sit well with people who were used to broadcast LANs, in which you plug a device in and it automatically learns where everything is. The newer ATM equipment can do this as well, thanks to PNNI and other autoconfiguration features.
"You still have to do the LANE configurations and set the network up to run IP, but it's getting much better," says Hans Baartmans, a Unix network administrator at Texas Instruments, Inc. in Dallas, who has been working with ATM off and on for seven years.
Moves and changes are also very easy. "The FORE 3810 is the simplest system I've ever had to configure," Chesapeake Energy's Sagebiel says. "I plug the cards in, and it knows how to configure the VLAN. If I want to go from 155M to 622M bit/sec, I just change the cards and quadruple the bandwidth over the same fiber."
Chesapeake's mission-critical campus network has dual redundant paths to all the buildings. "You can't do this with Ethernet - spanning tree shuts one down," Sagebiel says. "But ATM inherently likes redundant paths and does load balancing across them. The failover of services is extremely fast."
The ATM camp is hoping there will be some price/performance improvements in the fourth generation of ATM switches that is being developed.
"There are more silicon vendors working on chipsets, so they should be cheaper," says Ashok Madanahalli, product-line manager for ATM internetworking products at FORE in Pittsburgh. "OC-3 to the desktop starts at $400 per port now. And Microsoft's Windows 98 is going to ship with FORE drivers."
That sounds like a lot to Ethernet users, but it's actually a bargain if you do the economic analysis right, says Phu Dang, manager of computing solutions for Shell Exploration & Production Technology Co. (EPTC) in Houston.
Shell EPTC installed a FORE ATM network a year and a half ago to support more than 400 engineers and scientists engaged in energy research. It has an OC-12 backbone and OC-3 to all of the Unix workstations and some of the PCs. PCs that can't support 155M have to make do with switched 100M bit/sec Ethernet instead.
"The hardware is only 20% of the total cost," Phu says. "Things like troubleshooting and upgrading - that's the real cost. If high-speed networking is essential to what you do, ATM is the obvious solution."
ATM's future on the LAN hinges on the changing application mix and price points. Basic business applications today don't justify the cost. But that equation could change if time-sensitive applications start to proliferate and ATM prices drop enough.
Whatever happens, ATM has made a major contribution to the art of computer networking. It has focused people on the problem of QoS and on the notion that the network needs to help applications more, not just transmit data.
Winners and losers
ATM has a lot in common with token ring. Both are connection-oriented, offer superior technology and are getting clobbered by the brute force of Ethernet. But we aren't likely to see gigabit token-ring switches comprising enterprise backbones.Routing vs. bridging: an inverse relationship |
For one thing, ATM and token-passing FDDI have complemented token-ring LANs nicely, reducing the need to develop High-Speed Token Ring technology. Also, token ring isn't a contention technology like Ethernet, so there are no collisions that can be eliminated by switching to the desktop. Consequently, token-ring switches have been used largely to aggregate hubs. However, switched-port prices ($250 to $275) are now getting close enough to shared-port prices ($200) that switching to the desktop is practical.
"Switches are the one token-ring investment that continues to make sense," says Michelle McLean, senior analyst for META Group, Inc. in Burlingame, Calif.
There are no Layer 3 token-ring switches to date, partly because the existing switches function rather like Layer 2.5 devices. Token ring's source-route bridging avoids spanning tree problems by allowing for parallel paths in a Layer 2 network. As a result, there is less need for routing - whether traditional or via Layer 3 switches - in token-ring LANs.
"We also don't have as much of a reversal of the 80/20 rule on token-ring networks," says David Olechovsky, chairman of the High-Speed Token Ring Alliance (HSTRA) and token-ring product-line manager at IBM in Research Triangle Park, N.C.
Still reeling from the blow of Cisco's desertion, HSTRA members are regrouping this week at Interop with demonstrations of the first 4/16/100M bit/sec token-ring network interface cards (NIC). At press time, IBM was also planning to show off the first 100M bit/sec token-ring switch.
The next Cisco?
The move from traditional routers to routing switches is the kind of technological upheaval that creates opportunities for newcomers and threatens the established players. A lot of the best technology is coming from start-ups that hope the move to switching is a big enough paradigm shift to establish a new order.However, analysts are unanimous on this one: Not a chance. The move from shared to switched Ethernet didn't produce any new leaders, and the current generation of start-ups is fated to get acquired - possibly by telcos.
But the start-ups, as always, are responsible for much of the innovation.
"The big guys are tied to their installed bases and existing products," says Foundry's Demopoulos. "We will always be nine to 12 months ahead of them. There are companies out there that have mission-critical networks and can't wait for the big guys. Those people are our customers."
Incyte Pharmaceuticals, Inc. in Palo Alto, Calif., is a case in point. In 1997, the company's Ethernet network was choking under the load created by genomics processing and other compute-intensive scientific applications. Meeting 1998 business goals was expected to triple the traffic to an average of 800 gigabits per day.
"We were a big Bay shop before, but they couldn't meet our needs at the time," says Philip Kwan, manager of network operations and planning for Incyte. Instead, the company upgraded its backbone with Foundry gigabit switches that provide a multi-building facility with access to supercomputers and big server farms.
Over the horizon
A lot of piloting is going on this year on similar projects, and analysts expect backbone upgrades to go into full swing in 1999. "In two to three years, the core of the network will be routing-switch based, not router based," says Forrester's Hannigan.The new generation of products has pushed the envelope out enough that bandwidth and prices are not a big story anymore. "We're getting to the point where real value is not bigger and faster but more intelligent and more software-aware," says Fred McClimans, chairman and CEO of Current Analysis, Inc., a consultancy in Sterling, Va.
Instead, the focus should shift to incorporating services - directories and firewalls, for example - into the switches.
"The big emphasis will be on prioritization," McLean says. "Not for video or voice, but to make sure that your SAP traffic gets through no matter what else is going on."
Increasingly, management tools should be leveraging directory services. To date, network management has really been device management. "Now you can manage class of service and QoS and align management policies with the particular needs of your business environment," says Clint Ramsay, vice president of marketing for 3Com Corp.'s enterprise systems division in Boxborough, Mass.
In the Infonetics survey of high-speed LAN users, three-fourths of respondents said they would require QoS capabilities by 2000. However, most of them plan to provide it by over-provisioning their networks.
While QoS gets a lot of hype from vendors, it doesn't seem to be much more than a check-off item for users right now. "We're not looking at QoS, but rather at using policy-based management to control access to network resources," says Frank Bielecki, network manager at Sandia National Laboratories in Livermore, Calif.
A number of experts expect this functionality to be provided through Directory Enabled Networks (DEN).
"This is a radical new approach to managing bandwidth," says Sam Alunni, a vice president at Sterling Research, Inc. in Sterling, Mass. Policies are established in a single directory that switches can access. The switches provision network resources based on profiles set up for users and applications.
"With DENs, you start to get the same degree of control over frame-based networks that you used to have to go to ATM to get," Alunni says.
Four nines
Networks should also get more reliable as functions are moved into silicon and there are fewer parts. After all, the network component that fails most often is the routing software."We will finally see high availability for the masses," says LANNET's Bendori. As this approaches the 99.99% availability requirement of the Bellcore Network Equipment-Building System standard, it will enable business-quality voice over IP and finally unlock the tremendous potential of computer-telephony integration.
Layer 3 intelligence will get more democratic as well. It is expected to spread to the wiring closets and even out to the edges as WAN links go beyond E-1 speeds.
"Intelligence is getting cheap, so put it in the closet," says Douglas Hill, co-founder of Xylan Corp. in Calabasas, Calif.
When desktops are attached to Layer 3 switches, all sorts of possibilities open up. These include user-authentication capabilities that enable the network
to reconfigure itself based on the identity of the user; end-to-end QoS; and much more granular directory-enabled networking and intelligent multicast distribution.
The Resource Reservation Protocol (RSVP) feature is being incorporated into Microsoft's Windows desktop and needs to talk directly to a Layer 3-aware device if it is to be useful. It is difficult to implement end-to-end QoS via RSVP if there is a Layer 2 device between the Windows machine and the rest of the network.
Eliminating Layer 2?
Vinod Bhardwaj, the inventor of the first Ethernet switch, can see carrying things even further by putting a router at every port and eliminating Layer 2 altogether. Assuming the world converges on IP and silicon prices continue to fall, he thinks it could happen in three to four years."Routing is a superset of switching, so it has everything the network requires," says Bhardwaj, now president and CEO of ControlNet, Inc., a high-speed networking start-up in Campbell, Calif. "You would just need a NIC and a router."
Bhardwaj's vision isn't out of the realm of possibility. In fact, it dovetails nicely with the TCP/IP protocol stack, in which Layer 2 is equivalent to OSI's Layer 3 and there is no separate data-link layer.
First, though, the routers will have to get a lot more intelligent, says Mary Petrosky, senior analyst at The Burton Group in Salt Lake City. Otherwise, the configuration and administration of all those routers would be prohibitive.
"In ATM, if you have a network with some 155M switches and you drop a new 622M bit/sec box into the core, they will all find each other and update their tables and learn the new topology," Petrosky says. "Routers don't do that at this point."
Self-configuring routers that can fit into a single switch port at an affordable price? They're not exactly on the horizon, but with the speed at which switching has progressed thus far, who would say it can't happen?
Layer 4 switching: What it is and isn't
While Layer 2 and Layer 3 switches use brute force to speed up networks, Layer 4 switching attempts to add some finesse. That's the theory, anyway. In practice, the term has become another weapon for vendors waging a new round of "marketecture" wars.
Such weapons often are labeled with misnomers, and "Layer 4 switch" is no exception. In the seven-layer ISO model, packets are switched either to media access control addresses at Layer 2, the data-link layer, or to subnet addresses at Layer 3, the network layer. So-called Layer 4 switches merely look up into the transport layer of the packet to get information they can use to make smarter decisions about Layer 2 and Layer 3 forwarding.
For example, applications communicate with network services via an object called a port ID number. These TCP and User Datagram Protocol (UDP) port numbers tell the switch what type of application is generating the traffic, and the switch can then map the packet classifications into service guarantees.
In short, packets are just packets at Layer 2 and Layer 3. At Layer 4, there is knowledge about the sequence that an individual packet is part of and the application that generated it.
Traditional routers have had this Layer 4 functionality for years, but it degrades their performance so much that Layer 4 is almost never used. Today, some vendors claim their multilayer routing switches can process Layer 4 information and maintain wire-speed forwarding.
Moving further up the stack enables quality of service (QoS) and policy-based network management so administrators can fine-tune the use of the network with firewall-type granularity. For example, SAP R/3 traffic might be given priority over Web traffic, and bandwidth could be reserved for time-sensitive applications such as voice and video. And security policies can be applied at a much higher level, so hackers can't get in by just finding an IP address.
"What this means is that networks will become more services-oriented, and not just infrastructures for forwarding data," says Mary Petrosky, senior analyst with The Burton Group in Salt Lake City. The services that switches can support will depend on their ability to identify applications, which in turn "is what will separate the various vendors that are making claims about their Layer 4 products," she says.
Primitive, stateless applications such as telnet and File Transfer Protocol transmit on well-known TCP or UDP ports and are easy to spot. The same is not true, however, for the applications that really need prioritization, such as voice or enterprise resource planning software. These applications are state-dependent and don't have predefined port identifiers. Rather, the numbers get assigned dynamically by middleware, and the switch has to watch the sessions being established.
"You have exactly the same problem in Ethernet and ATM," says Donal Byrne, vice president of marketing for FDDI switch pioneer Berkeley Networks, Inc. in Milpitas, Calif. Switch manufacturers "don't do enough at Layer 4 to make their products useful to these stateful applications, which are the prominent and important applications on today's networks."
Berkeley Networks is tackling this problem by embedding Microsoft Corp.'s Windows NT operating system in its Gigabit Ethernet switch. This gives the platform access to all of NT's built-in services, including the directory, and creates what Byrne calls an "application-aware switch."
"We can take the thousands of network-based NT applications and services and integrate them on top of our platform according to the needs of our customers and partners," Byrne says. A separate policy server is not necessary. NT provides the translations between the applications and the hardware - a process that can take place at a relatively slow rate as long as the switch is doing the forwarding in hardware at wire speed.
"It's a neat idea, and it may enable Berkeley to implement policy-based networking before any other switch
manufacturer," says David Passmore, president of NetReference, Inc., a consultancy in Sterling, Va. Meanwhile, the big network companies are trying to lock customers in with announcements of their own architectures.
But all that's in the future. For now, Layer 4 switching seems to be a solution looking for a problem.
"Policy-based management? Most network administrators are still racing around on jet-powered skates troubleshooting," says Lynn DeNoia, director of consulting services for Strategic Networks in Rockland, Mass. "And people who are good at troubleshooting are not necessarily good at seeing things in a larger context and figuring out appropriate policies."
Where we've come from
New technologies tend to issue forth from agile start-ups unencumbered by installed bases and investments in existing product lines. LAN switching is a case in point. The first commercial Ethernet switch was prototyped in the Silicon Valley garage of entrepreneur Vinod Bhardwaj, now president and CEO of ControlNet, Inc., a high-speed net-working start-up in Campbell, Calif.
Flush from his stake in the initial public offering of former employer Excelan, Inc. (the TCP/IP specialist subsequently acquired by Novell, Inc.), Bhardwaj went out on his own in 1987 with an idea for boosting the capacity of what was then pre-10Base-T Ethernet. LANs were proliferating everywhere. Ethernet's bus architecture was holding things back.
Bhardwaj was working on a three-port device that would replace Ethernet's individual t-connectors when he had a flash: It was not the speed but rather the shared nature of Ethernet that was the problem. The three-port device simplified wiring but wouldn't really scale. The answer was a product that provided dedicated connections to each station and could eventually include an uplink to higher speed backbones.
When Bhardwaj pitched his invention to network companies, he was regarded, like many pioneers before him, as being afflicted with moonstruck madness. "They said, 'We've moved on to routing, and you're sending us back to bridging,' " he recalls.
Rejected but resolute, Bhardwaj left the established players to their 10Base-T committee battles and co-founded his own company with Larry Blair, now vice president of marketing for Redback Networks, Inc. in San Jose. Calif. They named the company after Bhardwaj's wife, Kalpana, whose name means "imagination" in Hindi.
The first order of business was to get rid of the "b-word," so the device - really the aggregation of a bunch of bridges - was designated an Ethernet switch. The product, dubbed the EtherSwitch, was encased in a traditional four-cornered box, but Kalpana marketers represented it on network diagrams as circular just to make it look different.
"FDDI and ATM were coming out at the same time that we were releasing this 'fancy bridge,' so we promoted it as just tactical," Blair says. Tactical indeed. It turned out to be the beginning of a paradigm shift that would enable frame-based networking to flourish into the 21st century. At the time, however, finding believers wasn't easy.
One well-known network industry executive was offered worldwide marketing rights to the EtherSwitch for $250,000. "But he turned us down three times," Blair recalls.
The Kalpana team was left to its own devices and in the summer of 1990 rolled out the first EtherSwitch. It had seven 10M bit/sec ports and sold for $11,500, about $1,650 per port. However, the price was less of an obstacle to prospective customers than the configuration changes to their networks.
"They were concerned about reliability and introducing a single point of failure," Bhardwaj remembers. "But once we got in the door, we could demonstrate a very visible improvement to network performance."
In fact, early product reviews characterized the EtherSwitch's speed as "stunning," and sales started to snowball. By 1992, vendors that had scoffed at the concept of Ethernet switching were lining up for OEM deals. Their customers were demanding "Kalpana-like" technology.
Kalpana passed into history in 1994 when it was swallowed up by router giant Cisco Systems, Inc. Ever the entrepreneur, Bhardwaj had left the company three months previously so he wouldn't have to sign a noncompete agreement.
What does he think about the revolution his EtherSwitch has wrought?
"Switching has progressed more than I originally thought, but it has also gotten a lot more complicated," he says. "Switches were throughput devices that were supposed to replace hubs, not routers. You keep them simple, simple, simple. Throw bandwidth at the problem, not complexity."
Switching grows up: Where we're going
How much faster and cheaper can data switching get? In eight years, switch throughput has gone from 150,000 packet/sec at Layer 2 to more than 50 million packet/sec at Layer 3, and there is no sign that the electronics driving these advances are running out of steam. In fact, recent announcements of terabit-speed switches indicate the rate of improvement may be accelerating.
Every time it starts to look as if the industry might have to go to optical technology in order to increase capacity, electronics makes another leap. The new terabit switches don't even incorporate the latest silicon technology. They use .25-micron silicon -1/600th the width of a human hair - and .18-micron technology is in the works. Such linear decreases in size translate into a geometric progression in the number of circuits that can be squeezed onto a chip. And the closer together the circuits, the faster they can operate.
"I tend to think there isn't a limit," says Diane Myers, a senior analyst who follows the semiconductor industry for In-Stat, a consultancy in Scottsdale, Ariz.
Six or seven years ago, Application Specific Integrated Circuits ran at 20 MHz and had 20,000 gates, or groups of transistors that implement some logic. Today, .25-micron technology has pushed those numbers to 100 MHz and 500,000 gates, and .15-micron silicon should bump them to 400 MHz and two million gates.
This is assuming chip designers use standard libraries that have been developed at the gate level. While using such prepackaged logic is a lot cheaper than starting from scratch, it wastes a lot of space.
"If they took the design down to the transistor level, they could do 400 MHz at .25 micron," says Donal Byrne, vice president of marketing for Berkeley Networks, Inc. in Milpitas, Calif.
The capacity of the basic materials is just one aspect of switch performance. Throughput can also be boosted dramatically through architectural innovations.
For example, typical shared-memory switches have trouble scaling beyond 30G bit/sec because the ports have to access the memory through a bus. Packet Engines, Inc., of Spokane, Wash., has eliminated the bus and come up with what it calls parallel access shared memory.
"In a multicast, we put each packet in shared memory, and any port that is supposed to take it does so," says Jeff White, vice president of marketing for Packet Engines. "There is no risk of oversubscribing the backplane - a typical problem for crossbar architectures, where you have to replicate all the packets in a multicast."
Power X, Ltd., a semiconductor start-up in Manchester, England, is addressing this problem and other crossbar limitations with a serial crossbar chipset that has separate matrix, control and application-interface modules. The first set, scheduled for release this September, can be used to build 80G bit/sec switches.
"We have developed a scheduling and arbitration mechanism to eliminate the head-of-line blocking associated with crossbar architectures," says Russell Johnson, vice president of sales and marketing for Power X. He expects the technology to support switch speeds of 320G bit/sec next year using existing software and to scale to the terabit range with more advanced software that's in the works. The Power X chipsets can be used in ATM and Ethernet switches.
If and when the electronics wizards run out of tricks, optical technology presents some possibilities. A 1988 patent describes a 30G bit/sec photonic-array backplane. But that technology emanated from the defense industry, in which checkbooks tend to be a bit larger than the ones available to Gigabit Ethernet start-ups.
"Affordable photonics technology for intelligent switching doesn't exist yet and probably won't for at least 10 years," says Mukesh Chatter, president and CEO of Nexabit Networks, a Westborough, Mass., start-up that recently announced a 6.4 terabit/sec switch aimed at the service-provider market.
One problem is that while electronics have memory, there is no method of storage for photonics yet. "You can do simple switching with devices that use mirrors to move the beams from one path to another, but that's a far cry from true optical routing," says Joe Ferguson, director of marketing for start-up Juniper Networks, Inc., a Nexabit competitor in Mountain View, Calif.
RELATED LINKS
Apply for your free subscription to Network World. Click here. Or get Network World delivered in PDF each week.
Request a reprint or permission to use this article.
