gabit Ethernet standard from the IEEE, 802.3z, will likely be ready late this year, with standards-confor
ming hardware arriving early next year.
But there is such demand for increased bandwidth that nonstandard gigabit-speed Ethernet products will begin to ship even before the IEEE finishes work on 802.3z. Gigabit Ethernet is gaining so much momentum that even companies that once talked about fielding competitive technologies -- such as Hewlett-Packard, with its Gigabit version of VG-AnyLAN -- are acquiescing.
But Gigabit Ethernet isn't a technology that companies can be cavalier about adopting. Vendors who see a business opportunity might be quick to jump on the bandwagon, but potential end users need to wrestle with infrastructure problems, not the least of which is Gigabit Ethernet's initial requirement for fiber connections.
Is Gigabit Ethernet in your company's future? To answer that, consider the following benefits and problems.
Backbones First
The first step in evaluating Gigabit Ethernet's potential is to understand how companies will initially
use it. The technology will first arrive in corporate backbones, either to increase total bandwidth or to replace multiple 100-Mbps networking links.
Market researcher Dataquest predicts that only some new killer asynchronous transfer mode (ATM) application could keep Gigabit Ethernet from becoming the dominant choice for backbone upgrades. By the year 2000, according to Dataquest, Gigabit Ethernet sales will reach $2.9 billion worldwide. By comparison, the company believes that ATM sales in 2000 will total just $1.5 billion, primarily for backbone applications. The technology might later establish itself for servers, especially Web servers and big data warehouses that have to be shared among a large number of users who will connect directly to a backbone switch using Gigabit Ethernet.
Given these applications, initial buyers will be mostly large companies, since they tend to have the most traffic to aggregate on their backbones. But some smaller users with data-intensive applications might also p
urchase Gigabit Ethernet. In either case, price/performance advantages will be the driving forces for Gigabit Ethernet.
Gigabit Ethernet switches, which could begin to appear around mid-year, will probably cost approximately $2200 per port, according to the Dell'Oro Group, a market-research and consulting firm in Portola Valley, California. Although they're pricey, these switches will have a significant performance advantage over 100-Mbps Ethernet and over 155-Mbps ATM, which will remain the predominant ATM speed this year (see the table
"What Gigabit Speed Will Cost"
for comparative prices). But as the pricing table shows, Gigabit Ethernet's price/performance advantage over both alternatives might improve steadily through the decade.
Gigabit vs. Fast Ethernet
Looking at the pricing estimates, you might conclude that 10 times the bandwidth for three to four times the cost will make Gigabit Ethernet a shoo-in for replacing multiple Fast Ethernet links. Howeve
r, the Dell'Oro Group's numbers cover only equipment costs. Cabling is another important consideration.
The initial Gigabit Ethernet standard calls for Ethernet frames over a Fibre Channel physical layer. Fibre Channel supports only multimode fiber, which can support Gigabit Ethernet runs of perhaps 300 to 500 meters, or single-mode fiber for distances of maybe 2 to 10 kilometers.
The cost of installing fiber-optic cabling, if it's not already in place, will raise the overall price of a Gigabit Ethernet installation. The cost of fiber-optic transceivers is also the major reason why Gigabit switches and network interface cards (NICs) are more expensive than their 10- and 100-Mbps counterparts, which are commonly UTP-based.
The requirement for fiber-optic cabling is least onerous in a backbone, where connections are few and long distances often require fiber in any case. For instance, when Hi Temp Insulation, a small manufacturing company located in Camarillo, California, had to connect three
of its buildings, it considered running fiber under the street for even a 100-Mbps connection. A consultant working with Hi Temp suggested that the company just go with Gigabit. The company is now installing MegaSwitch II switches from NBase Communications, which include prestandard Gigabit Ethernet modules.
Fiber might also be acceptable for centrally located servers, where the connection to the switch might be nothing more than a patch cord in a wiring closet. However, fiber is, by and large, a financial killer for desktop connections, which are both numerous and scattered.
For Gigabit Ethernet to become an affordable desktop solution, it must support UTP. Unfortunately, a UTP standard for Gigabit Ethernet could trail the initial fiber standard by a year or more. When a UTP standard does emerge, it will almost certainly employ multiple pairs to carry the Gigabit traffic. There's no technology on the horizon that can deliver Gigabit speeds over a single UTP cable pair.
The goal of 802.3z is
to provide full gigabit rates over the installed base of 100-meter UTP Category 5 cable by 1998. Several companies, including Broadcom, Lucent, National Semiconductor, Pacific-Sierra, and Silicon Design Labs, are working on proposals to achieve this goal. The problem is thorny, however, and no one knows yet how the UTP goal will be reached. In the meantime, some multiple-pair cable strategies are emerging; for more information, see the sidebar "Wire for a Gigabit World".
Gigabit Ethernet vs. ATM
By 1998, Gigabit Ethernet's main competition will come from ATM, which also excels in backbone applications. Compared to ATM, Gigabit Ethernet will offer more bandwidth for less money -- a compelling combination. Furthermore, since ATM also requires fiber, Gigabit Ethernet is at no disadvantage there.
In addition, with Gigabit Ethernet, LAN emulation isn't a concern, although it can be one of the more troublesome aspects of ATM. In fact, for companies already using Ethernet, Gigabit Eth
ernet requires essentially no retooling of the network management infrastructure or retraining of network administrators. That means savings of both time and money. In contrast, ATM requires new tools and retraining.
This is not to say that ATM is dead. First of all, ATM vendors will redouble their efforts to achieve speeds above 1 Gbps. ATM tops out at 622 Mbps today; however, it has been doubling its top speed every year. Continuing at that rate, ATM should reach 10 Gbps by the year 2000. Of course, by that time we might have 10-Gbps Ethernet, too. Although there may be some leapfrogging, ATM will probably always be faster -- but more expensive -- than Gigabit Ethernet.
In addition, ATM offers some potentially significant technical advantages. Specifically, ATM's small (53-byte) cells facilitate a fine-grained mixing of different traffic streams. Because ATM cells never vary in size, they allow a smoother mix of traffic streams.
Ethernet frames, in contrast, vary in size from 64 bytes to a
bout 1500 bytes, resulting in a much grainier and less predictable mix of traffic streams. Smooth, consistent traffic flow can be important for real-time, delay-sensitive traffic, such as video and voice. ATM is also ahead of Ethernet in defining quality of service (QoS) standards that make it possible for switches from different vendors to cooperate in handling traffic streams in ways that suit a particular type of traffic.
However, the majority of LAN segments today don't carry critical real-time video or voice streams, making ATM's strengths largely irrelevant in those environments. In addition, delivery can be quite smooth with Ethernet if bandwidth is plentiful in comparison to demand. "If you have infinite bandwidth, you don't need to manage it," Andy Bechtolsheim, vice president of Engineering for Cisco System's Gigabit switching group, points out.
Finally, switched Ethernet makes the mixing of multiple streams less of an is-sue by giving every pair of communicating nodes a private, clear c
hannel. Nearly all Gigabit Ethernet products will be switched. Gigabit Ethernet proponents are working on a QoS scheme as well.
Even in a switched environment, the backbone tends to mix many different traffic streams. Thus, ATM might win out over Gigabit Ethernet for backbones that mix data, video, and voice traffic, particularly over WANs.
Collision Domains
The issue of whether to support collision detection has been a sometimes-contentious one for the 802.3z study group. It's an integral and basic feature of the 802.3z standard. Unfortunately, collision detection and Gigabit Ethernet are a difficult match.
In collision detection, each node "listens" before transmitting and waits for the channel to be idle before issuing a transmit. Unfortunately, two nodes sometimes decide to transmit at almost exactly the same time. In that case, their transmissions collide, and neither one can get through. Both nodes detect the collision, wait a variable length of time, and try again.
The farther away the nodes are from one another, the longer it takes their signals to reach each other, and the more likely it becomes that one node has begun retransmitting and the other can't yet hear it.
To keep collisions down to a manageable level, the existing Ethernet standard specifies that the worst-case round-trip delay of the network must be less than or equal to the transmission time of the shortest legal frame. In other words, the amount of data that can be wrecked by a collision should be no more than the shortest frame, and a station should never have to retransmit more than one frame due to a collision.
The faster the network, the more bits a transmitter can put out in a given period of time, and the shorter the transmission time of the smallest frame. Since the actual speed of electrical signals traveling along the wire does not vary, the collision domain for a fast network must be smaller than that of a slower one.
For instance, the collision domains for 10-Mbps and 100-Mbp
s Ethernet are approximately 2000 and 200 meters, respectively. Following that pattern, the collision domain for Gigabit Ethernet should be about 20 meters, which is too short to be generally useful.
In practice, almost all Gigabit Ethernet products will handle this problem by eliminating collision detection entirely. In full-duplex switched environments, this won't create any problems. Since there are only two stations on each segment, and each has its own clear channel, no collisions can occur. Despite this practical reality, collision detection will still be part of the Gigabit Ethernet standard.
"It was largely a political issue," explains a standards participant who prefers to remain anonymous. "Once you pay for the fiber and fiber transceiver, there's no advantage to running half-duplex. I don't know anybody that will build half-duplex equipment. But it would be political suicide to try to define another network standard that's not contained in the 802.3 standards body. HP tried it with 802.
12 and basically shot itself in the foot. It had to be Ethernet-compatible."
Just in case somebody does produce a collision-based Gigabit Ethernet product -- a half-duplex fan-out device for PCs is one possibility -- two techniques have been incorporated into the standard to make a collision-based environment more workable.
One technique,
carrier extension
, makes very short frames appear to be longer than they actually are to ensure that collisions can be reliably detected. Specifically, all frames of 512 bytes or less appear to be 512 bytes long. Increasing the smallest legal frame size increases the collision domain proportionally.
But carrier extension can also waste a lot of bandwidth, especially for very short frames. This is a serious problem, since most control frames are short, and control frames are normally transmitted every few data frames.
To compensate for this tendency, a second technique,
packet bursting
, lets nodes send multiple frames back to back, ap
plying carrier extension only to the first frame in the burst. A single burst is limited to about 3 KB to prevent one node from hogging the network.
Using these two techniques, it might be possible to extend the Gigabit Ethernet collision domain to 200 meters while maintaining a 30 percent to 40 percent network utilization for small frames -- and perhaps as much as 90 percent utilization for large frames, according to Moti Weizman, director of hardware engineering with NBase Communications (Chatsworth, CA).
Network Design
Three-tier designs will be the most common way of implementing Gigabit Ethernet (see the figure
"Ethernet in Three Tiers"
). The lowest tier might consist of 10-Mbps Ethernet switches with 100-Mbps uplinks, with the second tier consisting of 100-Mbps Ethernet switches with Gigabit Ethernet uplinks, and the highest tier made up of either pure Gigabit or ATM switches. Network administrators can attach servers at any level, depending on whether
they're accessed by a workgroup, a business unit, or the entire enterprise.
At each level, the switches have increasingly fast backplanes. Such a design handles typical 10-Mbps workstation connections with inexpensive switches and uses more expensive, high-capacity switches only at the higher levels, where enough traffic can be aggregated to justify the cost.
Workgroup switches can implement a partially blocking architecture, which assumes that only a given percentage of nodes will ever have to transmit at full speed at the same time. If by chance this assumption is violated, only a few nodes and a limited geographical area are likely to be affected. Higher-level switches can be nonblocking, ensuring that packets are never lost at higher levels, where more nodes and larger geographical areas are likely to be affected.
The three-tier design also provides multiple layers of flow control and buffering, both of which are important when lower- and higher-speed ports are present in the same switc
h. Flow control ensures that traffic from a fast port doesn't overwhelm a device on a slower port, causing the device to drop packets. Dropped packets necessitate retransmission, resulting in less efficient use of the network overall.
Unfortunately, standard flow-control mechanisms, such as those defined in 802.3z, shut the high-speed port off completely, interrupting transmissions to all lower-speed ports, even if only one lower-speed port becomes overwhelmed (as indicated by a full buffer). The greater the difference in speed between the ports, the more likely such shutdowns are to occur. For instance, a 10-Mbps port is more likely to be overwhelmed by a Gigabit port than by a 100-Mbps port.
A three-tier design can ensure that a Gigabit port and a 10-Mbps port never communicate directly. Thus, an overwhelmed 10-Mbps port can shut down a 100-Mbps port, but not a Gigabit port. Only an overwhelmed 100-Mbps port (or another Gigabit port) can shut down a Gigabit port. A multitier design allows the ne
twork to tap the brakes rather than slamming them on at every little obstacle.
Buffering handles very short bursts of high-speed data without packet loss, again increasing the overall network efficiency. With multiple layers, the system has more "give" than it has when you employ only a single layer.
Faster and Faster
Initially, Gigabit Ethernet will be most popular where fiber-optic cabling is most common, distances are longest, traffic quantities are greatest, and the number of connections are fewest. That means primarily on enterprise-wide backbones, secondarily for servers, and seldom at the desktop level.
However, as prices fall, as UTP standards for Gigabit Ethernet emerge, and as servers and workstations continue to get more powerful, Gigabit Ethernet will look increasingly attractive for servers and the desktop. No doubt that's when talk of 10-Gbps Ethernet -- which is already beginning -- will start getting translated into action.
The following are end-user, per-port prices for backbone and
LAN-segmentation switches. (Source: Dell'Oro Group)
Ethernet with an Attitude
