Gigabit Ethernet - The Potential Illusion of Speed

Gilbert Held

Over the past two decades, a number of different types of Ethernet networks were standardized. While retaining the Carrier Sense Multiple Access/Collision Detection (CSMA/CD) access scheme, each major "flavor" of Ethernet differed from a preceding version by an increased network operating rate. The original version of Ethernet, which was standardized by the Institute of Electrical and Electronic Engineers (IEEE) as the 802.3 standard, operates at 10 Mbps. Although a 10 Mbps local area network operating rate was well suited for an era when transmission primarily consisted of moving text-based data between stations on a network, the technology began to see its age as graphic-based applications gained popularity.

Technological developments in chip sets and advances that resulted in improved cabling made it possible to literally supercharge Ethernet and enhance its operating rate by a factor of ten. The resulting technology, which is referred to as Fast Ethernet, actually represents three methods for obtaining a 100 Mbps operating rate using the CSMA/CD access protocol. One method, referred to as 100Base-T, requires transmission on 3-wire pairs in comparison to the dual-wire pairs used by the IEEE 802.3 standard known as 10Base-T. A second Fast Ethernet technology, referred to as 100Base-T4, supports a 100 Mbps operating rate through the use of 4-wire pairs. A third Fast Ethernet, referred to as 100Base-FX, defines transmission at 100 Mbps over fiber-optic media.

Currently Fast Ethernet in all its flavors represents the most popular type of LAN being installed by industry, government, and academia. However, the growth in the use of graphic-based applications to include signatures incorporated into email documents, photographs and drawings inserted into word processing documents, and the literal explosion in Web surfing has resulted in the need for more speed. Perhaps anticipating the saturation of some Fast Ethernet networks, the IEEE emphasized development of another revision to Ethernet. This newest version of Ethernet retains its CSMA/CD access protocol but increases its operating rate an order of magnitude beyond Fast Ethernet.

The resulting standard, referred to as Gigabit Ethernet, is similar to Fast Ethernet in that it is designed to operate over several types of media to include single and multi-mode fiber, shielded balanced copper such as coax, and unshielded twisted pair. Although the use of fiber is relatively expensive, it enables full-duplex transmission and avoids consideration of the Ethernet collision window as there are no collisions on a full-duplex link. Similarly, the use of Gigabit switches that support full-duplex operations alleviates collision window problems but is also relatively expensive. The use of a shared-media Gigabit Ethernet network, while relatively low in cost, is literally too fast for its own good and requires a modification in the form of extending the length of certain frames. As discussed later in this article, the resulting effect of the extension is to degrade the performance of Gigabit Ethernet to the point where under certain situations its ability to transport data represents only a marginal improvement over Fast Ethernet.

Mathematical Oddities

Simple mathematics tells us that 100 is ten times 10, and 1000 is ten times 100. If we apply logic and simple mathematics to the operating of Ethernet, Fast Ethernet, and Gigabit Ethernet, we would expect the operating rate of each local area network to differ by a similar amount. Although the logic and simple mathematics are indeed true and do result in operating rates increasing by an order of magnitude from Ethernet to Fast Ethernet to Gigabit Ethernet, what can we assume about the ability of each network to transport data? Does Fast Ethernet provide ten times the throughput of Ethernet? Similarly, does Gigabit Ethernet provide ten times the throughput of Fast Ethernet? As explained shortly, the answer to the second question under certain networking situations is negative, and the reason for the answer has important considerations for network managers and LAN administrators considering Gigabit technology.

Let's examine the basic frame format of Ethernet as a starting point for determining the obtainable frame rate on different types of Ethernet networks. By first computing the frame rate on a 10 Mbps network, we can use the resulting computations as a base to examine the effect of increasing the LAN operating rate upon the effective data transfer capability. Note that the effective data transfer rate represents the quantity of data per unit time transferred over a network. Because all protocols have a degree of overhead in the form of headers and trailers wrapped around information formed into frames (layer 2) or packets (layer 3), the effective data rate is always lower than the operating rate of a transmission facility. Although most people only consider the LAN operating rate, the effective data transfer rate is a more important metric as it governs the ability of a network to transport data.

Frame Rate Computations

Figure 1 illustrates the basic format of an Ethernet frame when the frame is placed onto a network. Although a minimum frame length of 64 bytes is typically referenced, this number refers to the length of the frame prior to its placement onto a network. Once the latter occurs, eight bits are added to each frame for synchronization in the form of a 7-byte Preamble field and a 1-byte Start of Frame Delimiter field. This results in a minimum Ethernet frame length of 72 bytes.

In examining Figure 1, note that the Data field ranges from a minimum of 46 bytes to a maximum length of 1500 bytes. When a maximum length frame is formed, its length is 1518 bytes in the network adapter. However, when placed on the network, the 8 bytes worth of Preamble and Start of Frame Delimiter fields result in a maximum frame length of 1526 bytes. Thus, all Ethernet frames (other than those in error) range between 72 and 1526 bytes in length when flowing on a network.

Because the length of an Ethernet frame can vary between 72 and 1526 bytes, the frame rate obtainable on a network will vary. However, since the minimum and maximum frame lengths are known, we can compute the frame rate in terms of each metric, which will provide the maximum and minimum frame rates on a network. Because frame rate is inversely proportional to frame length, the maximum frame length will enable us to compute the minimum frame rate, while the minimum frame length allows us to compute the maximum frame rate.

At a 10Mbps operating rate, Ethernet requires a dead time of 9.6 sec between frames. The bit duration at a 10 Mbps operating rate is 1/107 or 100 nsec. Thus, we can compute the maximum number of frames per second for maximum length and minimum length frames. For example, consider the maximum length frame of 1526 bytes. Here the time per frame becomes:

9.6 sec +1526 bytes *8 bits/byte * 100 nsec/bit

which results in a frame time of 1.23 msec. Thus, in one second there can be a maximum of 1/1.23 msec or 812 maximum length frames, each capable of transporting 1500 bytes of data. This means that if all frames were maximum length frames, the effective data transfer capability would be:

812 frames/sec * 1500 bytes/frame * 8 bits/ byte or 9.744 Mbps

We can revise the preceding computations to determine the number of minimum length frames that can flow on a 10 Mbps Ethernet network. For a minimum length frame of 72 bytes, the time per frame is:

9.6 sec +72 bytes * 8 bits/byte * 100 nsec/bit

which results in a frame duration of 67.2x10-06 sec. Thus, in one second there can be a maximum of 1/67.2x10-06 or 14880 minimum length 72 byte frames, each capable of transporting between one and 46 data characters. If all frames were minimum length frames, the effective data transfer rate would range between:

14880 frames/sec *1 byte/frame *8 bits/byte or 119040 bps to

14880 frames/sec * 46 bytes/frame * 8 bits/byte or 5.48 Mbps

Based upon the preceding computations, a file transfer between two stations on an Ethernet network that completely fills each frame's data field to its maximum capacity of 1500 bytes will result in an effective data transfer of 9.744 Mbps, approaching the 10 Mbps LAN operating rate. However, as file transfers are replaced by interactive queries that result in a lesser number of data characters transported in each frame, the effective data transfer rate decreases. On a probably absurd level, if each frame transported one data character the effective data transfer rate would be reduced to 119040 bps. Even when each minimum length frame is filled with 46 data characters, the effective data transfer capacity is only slightly over half the networks 10 Mbps operating rate.

Fast Ethernet uses the same frame format as Ethernet but the dead time between frames and bit duration are one tenth Ethernet's 10 Mbps metrics. Thus, the frame rate for maximum and minimum length frames are ten times that of Ethernet. That is, Fast Ethernet supports a maximum of 8120 maximum length 1526 byte frames per second and a maximum of 148800 minimum length 72 byte frames per second. Similarly, the effective data transfer capability of Fast Ethernet is ten times that of Ethernet. Table 1 compares the frame rates and effective data transfer capability of Ethernet and Fast Ethernet.

Although we might reasonably expect Gigabit Ethernet to extend Fast Ethernet's frame rate and data transfer capability by a factor of ten, this is not the case under certain situations. Those situations involve the use of Gigabit Ethernet in a shared media environment when the Gigabit Ethernet basic frame length is less than 512 bytes.

Gigabit Constraints

The use of Gigabit in a shared media environment requires a station on the network to be able to hear any resulting collision on the frame it is transmitting before it completes the transmission of the frame. This means that the transmission of the next-to-last bit of a frame that results in a collision should allow the transmitting station to hear the collision voltage increase prior to the transmission of the last bit. Thus, the maximum allowable Ethernet cabling distance is limited by the bit duration associated with the network operating rate and the speed of electrons flowing on the network.

When Ethernet operates at 1 Gbps, the allowable cabling distance would normally have to be reduced from Fast Ethernet's 200-meter diameter to approximately 10 meters or 33 feet. This would result in a major restriction on the ability of Gigabit Ethernet to be effectively used in a shared media, half-duplex environment. To overcome this cabling limitation, a carrier extension scheme was proposed by Sun Microsystems, Inc. This scheme, which results in the extension of the time an Ethernet frame is on the wire, became part of the Gigabit Ethernet standard for half-duplex operations.

Under the Gigabit Ethernet carrier extension scheme, the IEEE standard requires a minimum length frame of 512 bytes to be formed for shared-media, half-duplex operations. This means that the resulting frame when placed onto the network must be a minimum of 520 bytes in length due to the addition of a 7-byte Preamble and 1-byte Start of Frame Delimiter fields mentioned previously.

Figure 2 illustrates the Gigabit Ethernet carrier extension scheme associated with ensuring the flow of extended minimum length frames. In examining Figure 2, note that the timing extension occurs after the end of the standard Ethernet frame. The actual carrier extension occurs in the form of special symbols that result in the occurrence of line transitions and inform other stations "listening" to the network that the wire is in use. The carrier extension extends each Gigabit frame time to guarantee a minimum 512-byte slot time (520 bytes on the network) for half-duplex operations. Note that the increase in the minimum length of frames do not change the contents of any frames. Instead, carrier extension technology only alters the time the frame is on the network. Thus, compatibility is maintained between the original Ethernet frame and frames flowing on half-duplex Gigabit Ethernet. It should also be noted that Gigabit Ethernet carrier extension technology is not applicable for non-shared media environments, such as transmission on fiber or workstation connections to full-duplex switch ports. This is because no collisions are possible on such network environments, alleviating the need to employ carrier extension technology to ensure each frame flows on the network for sufficient duration to enable the transmitting station to recognize the occurrence of a collision within a 200-meter diameter.

Although carrier extension technology enables the cable length of a half-duplex Gigabit Ethernet network to be extended to a 200-meter diameter, the extension is not without cost. The primary cost is one of additional overhead and lower data throughput, since extension symbols added to the end of short frames waste bandwidth.

To gain an appreciation for the way carrier extension technology wastes bandwidth, consider the requirement to transmit a 64-byte record. When using Ethernet or Fast Ethernet, the record would be encapsulated within 26 bytes of overhead, resulting in a frame length of 90 bytes. When using Gigabit Ethernet as the transport mechanism the minimum length frame must be 520 bytes when flowing on the network. Thus, the frame must be extended through the addition of 430 (520-90) carrier extension symbols.

To further complicate bandwidth utilization, when the amount of data to be transported by a frame is less than 46 bytes, nulls are added to the data field to produce a 64-byte minimum length (72 bytes on the wire) frame before extending the frame via carrier extension symbols. Thus, a simple query, such as "Mr Watson, I hear you" (consisting of 20 characters to include spaces between words) would be placed in the Data field, padded with 26 nulls under each version of Ethernet to form a minimum length frame. However, under Gigabit Ethernet, the frame would be extended further through the addition of 448 carrier extension symbols to obtain a minimum 512-byte slot time or 520 bytes when the frame flows on the network. In this example, the ratio between actual information transported and total characters transmitted changes from 20 per 72 on Ethernet and Fast Ethernet to 20 per 520 on Gigabit Ethernet!

Under Gigabit Ethernet, the minimum 64-byte slot time (72 bytes on the network) requires the use of 448 carrier extension symbols. To examine the effect upon the data transport capability of Gigabit Ethernet, we can compute the frame rate in a manner similar to prior computations. However, since the minimum length frame is 520 bytes we will use that value instead of 72 bytes to compute the maximum frame rate. In doing so, the dead time between frames becomes .096 sec, and the bit duration becomes 1 nsec. Thus, the time per minimum length frame becomes:

.096 sec +520 bytes * 8 bits/byte * 1 ns/bit or  4.256 x 10-06 sec