CSMA/CD Gigabit Style

Gigabit Ethernet vendors are convinced that their customers want only Gigabit Ethernet products that support full-duplex mode. Thus, you shouldn't be surprised to learn that no Gigabit Ethernet products support half-duplex mode.

Nevertheless, the Gigabit Ethernet standard includes half-duplex specifications. These half-duplex specifications hinge on an enhanced version of the Carrier Sense Multiple Access with Collision Detection (CSMA/CD) protocol. The CSMA/CD enhancement increases the carrier event time from the traditional 512 bits (64 bytes) to 512 bytes (4,096 bits). (The carrier event time is the minimum amount of time a transmitting station must stay on the wire.) The 802.3z Task Force made this enhancement so that half-duplex Gigabit Ethernet segments could maintain a reasonable network diameter of 200 meters. (The network diameter is the maximum distance between two stations within a collision domain.)

The question that arises is why the 802.3z Task Force had to increase the carrier event time to maintain a reasonable network diameter. You must understand the problem the task force faced before you can fully appreciate its solution.

THE PROBLEM

Using the CSMA/CD protocol, stations on an Ethernet network listen to the wire, transmit when the wire is free, detect collisions, and wait to retransmit. For this system to work, a station must be able to detect a collision before that station retransmits--otherwise you would have undetected collision after collision and, basically, a big mess.

The original Ethernet Task Force established several values for the CSMA/CD protocol to ensure that a transmitting station detects a collision before retransmitting. One of these values is called the round-trip group delay time. The round-trip group delay time is based on the time it takes a signal to leave a transmitting station, travel to the opposite end of the wire, and return to the transmitting station. The Ethernet Task Force established a value for this round-trip group delay time that allowed a maximum network diameter of 2.5 kilometers for a 10 Mbit/s network.

When the Fast Ethernet Task Force increased the 10 Mbit/s signaling speed to 100 Mbit/s, they also needed to increase the round-trip group delay time by the same amount to ensure that transmitting stations detected collisions before retransmitting. For various reasons, however, the Fast Ethernet Task Force did not want to change the CSMA/CD algorithm and, thus, did not want to change this preestablished value. Instead, the Fast Ethernet Task Force reduced the network diameter: The task force multiplied 10 Mbit/s by 10 to get 100 Mbit/s and, in parallel, divided the 2.5-kilometer network diameter by 10 to get roughly 250 meters, which turned out to be 205 meters in practice.

The Gigabit Ethernet Task Force faced a similar problem. Like the Fast Ethernet Task Force, the Gigabit Ethernet Task Force did not want to change the CSMA/CD algorithm and, thus, did not want to change the value of the round-trip group delay time. However, reducing the network diameter wasn't a viable option either. If the Gigabit Ethernet Task Force had divided the network diameter by 10 to reflect the increase in the signaling speed, the theoretical network diameter would have been only 25 meters. By the time you built a practical network, the network diameter would have been even less than 25 meters. Clearly, reducing the network diameter was not the solution.

THE SOLUTION

Instead, the Gigabit Ethernet Task Force extended the slot time--without affecting the minimum frame size. Sounds impossible? It might. Traditionally, the slot time and the minimum frame size are identical. With Gigabit Ethernet, however, the minimum frame size is 64 bytes, but the slot time is 512 bytes. The Gigabit Ethernet Task Force worked this magic with a solution called a carrier extension.

A carrier extension is a nondata signal that Gigabit Ethernet devices add to the data fields of frames that are less than 512 bytes. (See Figure 2 .) (8B/10B, the signal encoding scheme on which the 802.3z specification is based, was conveniently equipped with nondata signals.) A carrier extension thus ensures that every frame on a Gigabit Ethernet network, even frames that are only 64 bytes, stay on the wire for a minimum of 512 bytes.

Although extending small frames may have prevented potential collisions that would otherwise occur on a Gigabit Ethernet network with a 200-meter network diameter, a new problem was introduced: Carrier extensions slowed performance.

To offset this performance problem, the Gigabit Ethernet Task Force added an option called packet bursting to the CSMA/CD algorithm. Packet bursting enables a transmitting station to send more than one frame of less than 512 bytes during one transmission event. During this transmission event, the transmitting station can send up to 3,000 bytes worth of small frames--roughly the equivalent of two maximum-sized frames.

THE LOGIC

While the carrier extension feature and the packet bursting option work, they add complexity that developers would have to build into half-duplex Gigabit Ethernet products. The irony is that this complex solution was designed to support half-duplex mode, which, most likely, no one will ever use. "But we're talking about standards," says Brian MacLeod, director of Marketing at Packet Engines Inc., "not necessarily logic."