There's a good chance you already have a port switch installed. It's by far the most mature of these three technologies. A switching hub is a type of segmented Ethernet hub that can reassign, on the fly, any port to any segment. Like all segmented hubs, the switching hub represents a compromise between high-cost, high-performance switching and low-cost, low-performance shared connections. On a shared hub, all stations attached to the hub contend for a single network segment. A segmented hub reduces contention by dividing the hub into multiple segments. A port switch optimizes the segmented architecture with flexible port assignment. Finally, a switch gives every port its own segment, so there is no contention within the
switch (though there might be blocking if switch buffers fill).
Port switches have been available for three or four years. Now, because of the demands of today's networks, many observers believe port switching is the future of the hub. "If you don't have port switching, you don't play. It will be a given, a commodity in the managed hub market," says Nate Walker, Cisco Systems' product manager for Gigabit Ethernet.
"Probably even most unmanaged hubs with more than eight ports will do port switching," adds Bradford Winkler, vice president of sales and marketing for LANart, a manufacturer of Gigabit Ethernet switches and network interface cards (NICs).
Aside from reducing collisions, what's the big deal? To start with, try easier moves and changes. Instead of having to rewire a station to connect it to a different segment, the change can be made in software, without visiting the wiring closet.
To the administrator, the ability to assign ports to
segments
is like the
"virtual LAN" feature offered by many switches -- with none of the delay introduced by a switch. A virtual LAN gives the appearance of a single segment, but a port switch gives real single-segment performance.
Some port switches also feature automatic load balancing, in which the hub assigns ports to segments based on traffic. Automatic load balancing may be dynamic, taking place without any external command, or static, requiring the administrator to issue a command to initiate the load-balancing operation. Some port switches support both.
Automatic load balancing provides bandwidth on demand, a more cost-effective alternative to dedicating a switch port to each user. However, automatic load balancing also makes it impossible to control which segment any particular station is on. Thus, for instance, if you want to isolate sensitive servers on their own segment, you have to put them on a separate hub.
With static configuration, the situation is exactly reversed: You don't get bandwidth on de
mand, but you can use segments for isolation. Or you can, for example, increase efficiency by putting stations that exchange a lot of traffic on the same segment, or you can isolate equipment for testing or repair.
The Switching Hour
Although port switching may be the future of hubs,
some observers
see switches, not hubs, as the future of networking. "For customers who don't need a lot of bandwidth, port switching may work just fine," says John Armstrong, principal analyst for networking with research and consulting firm Dataquest (San Jose, CA). "In general, though, with the price of a dedicated, managed switch port from a mainstream vendor down to $100 for a 10-Mbps port, or $250 for a 100-Mbps port, one has to ask, 'Why go half way?' The real issue for most customers is not whether to go with port switching or a dedicated switch, but whether to go with a dedicated 10-Mbps or 100-Mbps switch port."
Other observers aren't so sure about that. Strictly on the price fr
ont, some port switches will fall to $25 per port by the end of 1997, says Steve Stange, a product manager with Transition Networks, maker of the StackMaster Pro SPS2000 port switch. In contrast, Armstrong says, switch prices may hold more or less stable, having already dropped rapidly in the past year.
But the big argument is over performance: Some hub proponents question the need for a switch at all. "I seldom see a big value in having a private 10-Mbps switched Ethernet connection for every workstation," says
Greg Glasgow
, a vice president at LAN Solutions (San Diego, CA), a systems integrator. "It sounds good, but it's too expensive, and there's a performance hit going through a switch."
Furthermore, if you've got 100 workstations hitting one server, the server is the bottleneck, and a switch probably won't do much to improve performance, says Rand Morimoto, president of reseller Inacom Oakland (Oakland, CA). Spread the traffic among multiple servers, and a switch might d
o you some good.
In addition, many older PCs can't benefit from anything more than a 10-Mbps shared connection, says Dave Hoppock, vice president for sales with Thibault Associates (Pleasant Hill, CA), a systems integrator with many clients that are small to medium-size companies. He notes that many of his customers are putting 10-/100-Mbps cards "into everything" but are not using the 100-Mbps capability because the machines aren't even stressing the 10-Mbps hub. (Inacom's Morimoto notes that the high price of 100-Mbps hubs has also delayed upgrades to 100 Mbps.)
In the end, it seems likely that port switches will find a niche based on both price and features. "I think we'll see a lot of designs where servers and maybe some power users have 100-Mbps full-duplex ports on switches," says Hoppock. "The rest of the users will be on 10-Mbps shared hubs or port switches."
Faster Routes
While hubs and switches alone can handle sizable workgroups, organizations with more than a few hundred
workstations typically also need routers. Routers segment the network to enhance performance, enforce security, and manage the flow of data according to the company's policies. Unfortunately, router often means "bottleneck": The route-computation engine has to extract information from each packet and make often-complex decisions based on it. As networks grow, routing tables grow also, and routing tends to become slower and slower.
Two basic approaches to this problem are evolving. Both approaches are most likely to be implemented in backbone switches and in distribution switches that sit between workgroup switches and backbone switches. Both approaches also focus on IP. Other Layer 3 protocols, such as IPX and AppleTalk, are bridged, not routed, so Layer 3 switching does nothing to improve their efficiency.
The first approach -- which has so far been used only with ATM, not with Ethernet -- reduces the results of route computation to a single piece of information, which is inserted into the packe
t and thereafter used by switching engines to determine the switching path. Because new protocols or protocol modifications are involved, upgrading one or two network devices does no good. Ideally, all switches and routers in the network should adopt the new protocol (see "Faster, Smarter Nets," April BYTE).
The second approach, pioneered by Rapid City Communications (acquired by Bay Networks in June 1997), centers around an ASIC that performs route computation at switch-like speeds. Rapid City implemented unmodified IP routing in an ASIC in its F1200 Gigabit Ethernet routing switch. The ASIC enables the F1200 to do IP routing at switching speeds, without any new or modified protocols. Each switch port has its own ASIC, so you can turn routing on for individual ports. Unlike approaches that depend on new protocols or protocol modifications, the second approach gives results even if you upgrade only one switch.
At Networld+Interop in May, the F1200 forwarded 7 million packets per second (pps). By c
omparison, high-end routers may forward 1.5 million pps, while high-end switches may hit 2 to 5 million pps. The more complex the routing task, the poorer the performance. For instance, a priority scheme can degrade overall routing performance -- although priority traffic might get better performance. With routing in an ASIC, the routing task doesn't affect performance.
Even Faster Ethernet
Which brings us to Gigabit Ethernet. The initial question with Gigabit Ethernet is where to install it first. A longer-term question is to what extent Gigabit Ethernet will be used instead of asynchronous transfer mode (ATM).
On the question of initial installation, it's natural to look to the backbone. "Traditionally, there has been a concept of a hierarchy of bandwidth in Ethernet network design," says Jeff Wilbur, director of hub products in the networking products division of Compaq Computer. "You might start out at the lowest level with 10 Mbps shared, then go to 10 Mbps switched one level up, the
n to 100 shared, and so on. Gigabit Ethernet fits in very naturally at the top of that pyramid."
Alteon Networks, a vendor targeting the server market with its Gigabit Ethernet products (the AceSwitch and AceNIC), argues that servers are the safer place to get your first production experience with Gigabit Ethernet: Fewer users are affected by a failure. You may even be able to limit Gigabit Ethernet to server-to-server traffic, such as replication, or to bulk data transfer operations, such as backup. Users would probably not know about any problems on the Gigabit links.
Gigabit Ethernet and ATM
In the end, however, Gigabit Ethernet will be important for backbones. Companies that have ATM backbones largely because of its scalability beyond 622 Mbps may now be drawn back into the Ethernet fold.
"The technology developed for Gigabit Ethernet provides a solid base for 10-Gbps Ethernet," notes David Cheriton, a professor in the department of computer science at Stanford University. ATM n
o longer has a scalability advantage, he says.
It's widely accepted that ATM can provide smoother delivery for real-time traffic such as video and voice. But, Cheriton points out, the maximum packet in Gigabit Ethernet lasts 12 microseconds. "Human beings don't even begin to notice delay and jitter for voice and video until they're at least in the tens if not in the hundreds of milliseconds," he says. Rather than trying to manage bandwidth carefully to avoid delay, high-speed, inexpensive, switched Ethernet technologies will allow enough bandwidth to handle voice and video smoothly without special management techniques, says Cheriton.
Once sites start deploying Gigabit Ethernet widely on their backbones, there are arguments for migrating to a pure Ethernet backbone over time. "The trouble with moving from Ethernet to ATM and back to Ethernet is that you've got to disassemble the Ethernet packet into ATM cells, and then reassemble the packet when it gets where it's going," says Neal Upton, presiden
t of LANTech, an Indianapolis-based reseller.
Since carriers do not offer Ethernet interfaces, corporations might maintain some ATM just to interface with the WAN. But Cheriton thinks even that could change over the long run. He believes that if corporate networks are dominated by Ethernet, carriers and service providers will be motivated to provide Ethernet interfaces for customers and eventually even to use Ethernet for their own long-haul links. "The transition issue is the need for more buffering in WAN switches and routers," says Cheriton. "Initial switches and routers may be more restricted in buffering than what you might want for the WAN."
He points out that ATM switches suffer from the same problem, though. "ATM switches are dramatically low on buffering, and studies show how badly the first generation of switches is working. High-speed memory is an expensive component. You have to have a little bit of pain before vendors are willing to step up to putting the right amount of memory in the
ir products."
Decisions, Decisions
The decision to use unmanaged hubs, segmented hubs/port switches, or switches at the workgroup level is one that comes up for every network design. Despite the widespread popularity of switches, many resellers think hubs can save their customers a few dollars without affecting performance, and maybe adding a few features to boot.
Layer 3 switching, on the other hand, will work only for large networks for now. The software-based approaches to Layer 3 switching will affect only large ATM networks -- and only a minority of those, at least until the competitive situation sorts itself out. Resellers should understand software-based Layer 3 switching, but mostly they can steer their customers away for the time being. Silicon-based Layer 3 switching is a less risky proposition. It can fit transparently into existing networks and provide impressive throughput while preserving the routing architecture, but it has a drawback: It does not help optimize currently in
stalled equipment.
Gigabit Ethernet is still in the pioneering stages but is moving fast. Small networks probably don't need it today, but many medium and large networks will incorporate it over the next year or so. Resellers should be helping their customers find the best initial uses for the technology as it follows the usual curve from cutting edge to commodity.
Where to Find
Alteon Networks
San Jose, CA
Phone: 888-258-3661
Phone: 408-360-5500
Internet:
http://www.alteon.com
A Major Switch in Network Design
