Fast Ethernet

Fast Ethernet

At the same time that switches were becoming popular, the speed of 10-Mbps Ethernet was coming under pressure. At first, 10 Mbps seemed like heaven, just as cable modems seemed like heaven to the users of telephone modems. But the novelty wore off quickly. As a kind of corollary to Parkinson’s Law (‘‘Work expands to fill the time available for its completion’’), it seemed that data expanded to fill the bandwidth available for their transmission.

Many installations needed more bandwidth and thus had numerous 10-Mbps LANs connected by a maze of repeaters, hubs, and switches, although to the network managers it sometimes felt that they were being held together by bubble gum and chicken wire. But even with Ethernet switches, the maximum bandwidth of a single computer was limited by the cable that connected it to the switch port.

It was in this environment that IEEE reconvened the 802.3 committee in 1992 with instructions to come up with a faster LAN. One proposal was to keep 802.3 exactly as it was, but just make it go faster. Another proposal was to redo it totally and give it lots of new features, such as real-time traffic and digitized voice, but just keep the old name (for marketing reasons). After some wrangling, the committee decided to keep 802.3 the way it was, and just make it go faster. This strategy would get the job done before the technology changed and avoid unforeseen problems with a brand new design. The new design would also be backwardcompatible with existing Ethernet LANs. The people behind the losing proposal did what any self-respecting computer-industry people would have done under these circumstances: they stomped off and formed their own committee and standardized their LAN anyway (eventually as 802.12). It flopped miserably.

The work was done quickly (by standards committees’ norms), and the result, 802.3u, was approved by IEEE in June 1995. Technically, 802.3u is not a new standard, but an addendum to the existing 802.3 standard (to emphasize its backward compatibility). This strategy is used a lot. Since practically everyone calls it fast Ethernet, rather than 802.3u, we will do that, too.

The basic idea behind fast Ethernet was simple: keep all the old frame formats, interfaces, and procedural rules, but reduce the bit time from 100 nsec to 10 nsec. Technically, it would have been possible to copy 10-Mbps classic Ethernet and still detect collisions on time by just reducing the maximum cable length by a factor of 10. However, the advantages of twisted-pair wiring were so overwhelming that fast Ethernet is based entirely on this design. Thus, all fast Ethernet systems use hubs and switches; multidrop cables with vampire taps or BNC connectors are not permitted.

Nevertheless, some choices still had to be made, the most important being which wire types to support. One contender was Category 3 twisted pair. The argument for it was that practically every office in the Western world had at least four Category 3 (or better) twisted pairs running from it to a telephone wiring closet within 100 meters. Sometimes two such cables existed. Thus, using Category 3 twisted pair would make it possible to wire up desktop computers using fast Ethernet without having to rewire the building, an enormous advantage for many organizations.

The main disadvantage of a Category 3 twisted pair is its inability to carry 100 Mbps over 100 meters, the maximum computer-to-hub distance specified for 10-Mbps hubs. In contrast, Category 5 twisted pair wiring can handle 100 m easily, and fiber can go much farther. The compromise chosen was to allow all three possibilities, as shown in Fig. 4-19, but to pep up the Category 3 solution to give it the additional carrying capacity needed.

The Category 3 UTP scheme, called 100Base-T4, used a signaling speed of 25 MHz, only 25% faster than standard Ethernet’s 20 MHz. (Remember that

Manchester encoding, discussed in Sec. 2.5, requires two clock periods for each of the 10 million bits sent each second.) However, to achieve the necessary bit rate, 100Base-T4 requires four twisted pairs. Of the four pairs, one is always to the hub, one is always from the hub, and the other two are switchable to the current transmission direction. To get 100 Mbps out of the three twisted pairs in the transmission direction, a fairly involved scheme is used on each twisted pair. It involves sending ternary digits with three different voltage levels. This scheme is not likely to win any prizes for elegance, and we will skip the details. However, since standard telephone wiring for decades has had four twisted pairs per cable, most offices are able to use the existing wiring plant. Of course, it means giving up your office telephone, but that is surely a small price to pay for faster email.

100Base-T4 fell by the wayside as many office buildings were rewired with Category 5 UTP for 100Base-TX Ethernet, which came to dominate the market. This design is simpler because the wires can handle clock rates of 125 MHz. Only two twisted pairs per station are used, one to the hub and one from it. Neither straight binary coding (i.e., NRZ) nor Manchester coding is used. Instead, the 4B/5B encoding we described in Sec 2.5 is used. 4 data bits are encoded as 5 signal bits and sent at 125 MHz to provide 100 Mbps. This scheme is simple but has sufficient transitions for synchronization and uses the bandwidth of the wire relatively well. The 100Base-TX system is full duplex; stations can transmit at 100 Mbps on one twisted pair and receive at 100 Mbps on another twisted pair at the same time. 

The last option, 100Base-FX, uses two strands of multimode fiber, one for each direction, so it, too, can run full duplex with 100 Mbps in each direction. In this setup, the distance between a station and the switch can be up to 2 km.

Fast Ethernet allows interconnection by either hubs or switches. To ensure that the CSMA/CD algorithm continues to work, the relationship between the minimum frame size and maximum cable length must be maintained as the network speed goes up from 10 Mbps to 100 Mbps. So, either the minimum frame size of 64 bytes must go up or the maximum cable length of 2500 m must come down, proportionally. The easy choice was for the maximum distance between any two stations to come down by a factor of 10, since a hub with 100-m cables falls within this new maximum already. However, 2-km 100Base-FX cables are too long to permit a 100-Mbps hub with the normal Ethernet collision algorithm. These cables must instead be connected to a switch and operate in a full-duplex mode so that there are no collisions.

Users quickly started to deploy fast Ethernet, but they were not about to throw away 10-Mbps Ethernet cards on older computers. As a consequence, virtually all fast Ethernet switches can handle a mix of 10-Mbps and 100-Mbps stations. To make upgrading easy, the standard itself provides a mechanism called autonegotiation that lets two stations automatically negotiate the optimum speed (10 or 100 Mbps) and duplexity (half or full). It works well most of the time but is known to lead to duplex mismatch problems when one end of the link autonegotiates but the other end does not and is set to full-duplex mode (Shalunov and Carlson, 2005). Most Ethernet products use this feature to configure themselves.

Gigabit Ethernet

The ink was barely dry on the fast Ethernet standard when the 802 committee began working on a yet faster Ethernet, quickly dubbed gigabit Ethernet. IEEE ratified the most popular form as 802.3ab in 1999. Below we will discuss some of the key features of gigabit Ethernet. More information is given by Spurgeon (2000).

The committee’s goals for gigabit Ethernet were essentially the same as the committee’s goals for fast Ethernet: increase performance tenfold while maintaining compatibility with all existing Ethernet standards. In particular, gigabit Ethernet had to offer unacknowledged datagram service with both unicast and broadcast, use the same 48-bit addressing scheme already in use, and maintain the same frame format, including the minimum and maximum frame sizes. The final standard met all these goals.

Like fast Ethernet, all configurations of gigabit Ethernet use point-to-point links. In the simplest configuration, illustrated in Fig. 4-20(a), two computers are directly connected to each other. The more common case, however, uses a switch or a hub connected to multiple computers and possibly additional switches or hubs, as shown in Fig. 4-20(b). In both configurations, each individual Ethernet cable has exactly two devices on it, no more and no fewer.

Also like fast Ethernet, gigabit Ethernet supports two different modes of operation: full-duplex mode and half-duplex mode. The ‘‘normal’’ mode is fullduplex mode, which allows traffic in both directions at the same time. This mode is used when there is a central switch connected to computers (or other switches) on the periphery. In this configuration, all lines are buffered so each computer and switch is free to send frames whenever it wants to. The sender does not have to sense the channel to see if anybody else is using it because contention is impossible. On the line between a computer and a switch, the computer is the only possible sender to the switch, and the transmission will succeed even if the switch is currently sending a frame to the computer (because the line is full duplex). Since

no contention is possible, the CSMA/CD protocol is not used, so the maximum length of the cable is determined by signal strength issues rather than by how long it takes for a noise burst to propagate back to the sender in the worst case. Switch\%es are free to mix and match speeds. Autonegotiation is supported just as in fast Ethernet, only now the choice is among 10, 100, and 1000 Mbps.

The other mode of operation, half-duplex, is used when the computers are connected to a hub rather than a switch. A hub does not buffer incoming frames. Instead, it electrically connects all the lines internally, simulating the multidrop cable used in classic Ethernet. In this mode, collisions are possible, so the standard CSMA/CD protocol is required. Because a 64-byte frame (the shortest allowed) can now be transmitted 100 times faster than in classic Ethernet, the maximum cable length must be 100 times less, or 25 meters, to maintain the essential property that the sender is still transmitting when the noise burst gets back to it, even in the worst case. With a 2500-meter-long cable, the sender of a 64-byte frame at 1 Gbps would be long finished before the frame got even a tenth of the way to the other end, let alone to the end and back.

This length restriction was painful enough that two features were added to the standard to increase the maximum cable length to 200 meters, which is probably enough for most offices. The first feature, called carrier extension, essentially tells the hardware to add its own padding after the normal frame to extend the frame to 512 bytes. Since this padding is added by the sending hardware and removed by the receiving hardware, the software is unaware of it, meaning that no changes are needed to existing software. The downside is that using 512 bytes worth of bandwidth to transmit 46 bytes of user data (the payload of a 64-byte frame) has a line efficiency of only 9%.

The second feature, called frame bursting, allows a sender to transmit a concatenated sequence of multiple frames in a single transmission. If the total burst is less than 512 bytes, the hardware pads it again. If enough frames are waiting for transmission, this scheme is very efficient and preferred over carrier extension.

In all fairness, it is hard to imagine an organization buying modern computers with gigabit Ethernet cards and then connecting them with an old-fashioned hub to simulate classic Ethernet with all its collisions. Gigabit Ethernet interfaces and switches used to be expensive, but their prices fell rapidly as sales volumes picked up. Still, backward compatibility is sacred in the computer industry, so the committee was required to put it in. Today, most computers ship with an Ethernet interface that is capable of 10-, 100-, and 1000-Mbps operation and compatible with all of them.

Gigabit Ethernet supports both copper and fiber cabling, as listed in Fig. 4-21. Signaling at or near 1 Gbps requires encoding and sending a bit every nanosecond. This trick was initially accomplished with short, shielded copper cables (the 1000Base-CX version) and optical fibers. For the optical fibers, two wavelengths are permitted and result in two different versions: 0.85 microns (short, for 1000Base-SX) and 1.3 microns (long, for 1000Base-LX).

Signaling at the short wavelength can be achieved with cheaper LEDs. It is used with multimode fiber and is useful for connections within a building, as it can run up to 500 m for 50-micron fiber. Signaling at the long wavelength requires more expensive lasers. On the other hand, when combined with singlemode (10-micron) fiber, the cable length can be up to 5 km. This limit allows long distance connections between buildings, such as for a campus backbone, as a dedicated point-to-point link. Later variations of the standard allowed even longer links over single-mode fiber.

To send bits over these versions of gigabit Ethernet, the 8B/10B encoding we described in Sec. 2.5 was borrowed from another networking technology called Fibre Channel. That scheme encodes 8 bits of data into 10-bit codewords that are sent over the wire or fiber, hence the name 8B/10B. The codewords were chosen so that they could be balanced (i.e., have the same number of 0s and 1s) with sufficient transitions for clock recovery. Sending the coded bits with NRZ requires a signaling bandwidth of 25% more than that required for the uncoded bits, a big improvement over the 100% expansion of Manchester coding.

However, all of these options required new copper or fiber cables to support the faster signaling. None of them made use of the large amount of Category 5 UTP that had been installed along with fast Ethernet. Within a year, 1000Base-T came along to fill this gap, and it has been the most popular form of gigabit Ethernet ever since. People apparently dislike rewiring their buildings.

More complicated signaling is needed to make Ethernet run at 1000 Mbps over Category 5 wires. To start, all four twisted pairs in the cable are used, and each pair is used in both directions at the same time by using digital signal processing to separate signals. Over each wire, five voltage levels that carry 2 bits are used for signaling at 125 Msymbols/sec. The mapping to produce the symbols from the bits is not straightforward. It involves scrambling, for transitions, followed by an error correcting code in which four values are embedded into five signal levels.

A speed of 1 Gbps is quite fast. For example, if a receiver is busy with some other task for even 1 msec and does not empty the input buffer on some line, up to 1953 frames may have accumulated in that gap. Also, when a computer on a gigabit Ethernet is shipping data down the line to a computer on a classic Ethernet, buffer overruns are very likely. As a consequence of these two observations, gigabit Ethernet supports flow control. The mechanism consists of one end sending a special control frame to the other end telling it to pause for some period of time. These PAUSE control frames are normal Ethernet frames containing a type of 0x8808. Pauses are given in units of the minimum frame time. For gigabit Ethernet, the time unit is 512 nsec, allowing for pauses as long as 33.6 msec.

There is one more extension that was introduced along with gigabit Ethernet. Jumbo frames allow for frames to be longer than 1500 bytes, usually up to 9 KB. This extension is proprietary. It is not recognized by the standard because if it is used then Ethernet is no longer compatible with earlier versions, but most vendors support it anyway. The rationale is that 1500 bytes is a short unit at gigabit speeds. By manipulating larger blocks of information, the frame rate can be decreased, along with the processing associated with it, such as interrupting the processor to say that a frame has arrived, or splitting up and recombining messages that were too long to fit in one Ethernet frame.

Frequently Asked Questions