Ethernet or IEEE 802.3

TOPOLOGY

Vendors provide Ethernet in three cabling styles: thick 50 Ohm coaxial cable (approximately 1.4 cm or 0.625 inch in diameter), thin 50 Ohm coaxial cable (approximately 0.5 cm or 0.25 inch in diameter), and twisted-pair (telephone wire). Thick Ethernet is deployed in segments up to 500 m or 1640.4 ft. See Figure 1. Each end of a segment must be terminated with a 50 Ohm resistor connected between the outer shield and center conductor. Stations "tap" into the passive broadcast medium with transceivers.

To create a tap, the network manager bores a small hole through the coaxial cable's outer insulation and shield. The network manager then clamps onto the cable a transceiver housing containing a thin probe which punctures the inner in- sulation and makes contact with the center conductor. This technique provides an electrical connection with the cable but with signal reflection and attenuation of less than one percent.

The transceiver translates 0 to 3.6V signals from the Ethernet interface into 0 to -2V signals which drive the 500m cable. The transceiver also detects corrupted Manchester en- coded bits and reports them to the Ethernet interface which passes the error onto the computer interface as a frame col- lision. The transceiver cable may be up to 50m or 164ft long. Up to 100 transceivers may be placed on one coaxial cable segment but, to minimize reflections, no closer than 2.5m or 8.2ft. Up to five coaxial cable segments (IEEE 802.3) may be connected in series via local and remote re- peaters. A local repeater interconnects two segments via transceiver cables which run to a transceiver on each seg- ment. Remote repeaters interconnect two segments via a high speed link (such as a fiber optic cable) and may be up to 1km or 3,281ft long.

Thin Ethernet is much like thick Ethernet except for the cable diameter and the lack of transceiver cabling. Cable diameter is a function of the center conductor's diameter relative to the shield diameter, thus a smaller center con- ductor allows thin Ethernet to be more flexible and less ex- pensive, yet keep an impedance of 50 Ohms. In thin Ethernet, transceivers are located within the host interface. The thin cable is run directly between computer systems but passively attaches to each system with a twist-lock BNC "T" connector. The BNC T connector limits the distance of the cable tap to several centimeters or approximately one inch, thus, reducing signal reflections. Shielded connector notwithstanding, a BNC connector is more invasive than a thick Ethernet transceiver connection. BNC connectors produce approximately three percent signal attenuation per connector-barrel-pair. Cable length between systems must be at least 0.5m or 1.6ft, again, to reduce signal refections. Total thin cable length is limited to 185m or 606.9ft and 30 BNC station taps.

SIGNAL REFLECTION

Signal reflection effects many transmission systems and is a dominant factor high frequency electrical transmission systems. Simply stated, reflection is the energy of trans- mission systems lost to incompatibilities between the trans- mission medium and "receiver." For example, sound broadcast in a deep canyon reflects the transmission in the form of an echo. The sound is completely canceled at the interface be- tween air and sheer rock canyon wall and reflected back to the transmitter. A less dramatic example occurs at night when a flashlight is shined through a glass window into a dark room. More than 50% of the light is reflected back to the observer while the remaining signal passes through the glass. As as an opposite example, electrical power distribu- tion systems are one of the few cases where signal reflec- tions are not a design issue. In typical household electri- cal circuits "power hungry" appliances gobble up the trans- mitted signal (60Hz AC) plus any stray reflections which hap- pen by the power receptacle.

In high frequency broadcast systems such as Ethernet signal reflection plays a major role in signal attenuation. The broadcast medium, a segment of coaxial cable, is termi- nated at each end with 50 Ohm resistors. Ideally, the signal leaves the transmitter and propagates down the cable from the transmitter in both directions until it reaches the resistors at the end of the cable which dissipate the electrical signal in the form of heat. In real life, the signal propagates down the cable loosing strength as it fights the resistance of the conductor and magnetic fields (impedance) until it hits an obstacle (the metal probe of a transceiver or a BNC connector). A small amount of signal energy is reflected back to the transmitting station without interfering with the main signal reception at those earlier stations. With each obstacle encountered the signal is attenuated by approximate- ly one percent on thick Ethernet or three percent on thin Ethernet (EEs refer to this as an impedance mismatch). Therefore, in a fully configured Ethernet (thick or thin) ap- proximately one percent of the signal remains when the signal travels from a transmitter at one end of the cable to a re- ceiver at the other end.

But signal attenuation is only half the problem with re- flected signals. The reflected signal itself can combine with subsequent transmitted signals in different ways. If a reflected signal is present when a transmitted signal is pre- sent, amplification occurs. If a reflected signal is not present when a transmitted signal is present, cancellation occurs (i.e., the receiver sees no change in voltage). If a reflected signal is present a fraction of the time when a transmitted signal is present the transmitted signal may shift frequencies, be amplified, or attenuated.

LAN managers must be aware of the effects of signal re- flection and NOT fall into common pitfalls. Exact station numbers and exact cable run length between stations are rarely maintained. Thus, when installing transceivers or BNC "T" connections up to the "limit" and/or running cable up to the "limit" WILL lead to excessive signal attenuation and in- termittent frame transmission. Note that just having a BNC "T" connector on the cable has the same effect as having a station connected to the Ethernet. A second pitfall is the installation of cable "jumpers" less than "modulo minimum station distance," thus insuring that reflections will dis- tort transmitted signals and that stations will become unre- liable. A third pitfall is the placement of thin Ethernet BNC "T" connectors in the ceiling and installing a "tail" ca- ble to the station thereby introducing signal reflections and excessive signal attenuation. To be more specific, the sig- nal splits in the ceiling with the attenuated portion going down the main cable while the remainder travels to the com- puter and high impedance Ethernet receiver. The unused sig- nal returns to the ceiling and propagates in both directions on the main cable. Finally, a break in thick Ethernet or disconnecting a thin Ethernet BNC connector will "crash" the LAN on either side of the break because of signal reflections from the unterminated cable ends back into the two new seg- ments.

TOPOLOGY REVISITED

Many LAN managers remedy signal reflection pitfalls with local repeaters which create multiple electrically indepen- dent Ethernet segments. Thus a break in one segment does not effect other segments. The LAN manager then arranges Ether- net segments based on convenience and availability. Conve- nience dictates the use of a central wiring closet to provide quick identification of failed segments and repeaters as well as provide quick routing of new segments. Availability dic- tates that critical central file, boot, and mail servers be placed on their own hidden segment and thereby allow server access for all nondamaged Ethernet segments.

The LAN manager may also employ a bridge/repeater to create Ethernet segments and improve network performance. Bridges do not forward Ethernet frames if the addressee does not exist on the bridged segment. Thus if to two or more computers which share file systems are placed on one segment, then other segments will NOT see files exchanged among the computers. Finally, LAN managers can improve network perfor- mance by creating parallel segments among high volume ma- chines and then running each segment to a separate set of users. This last technique, however, requires the protocol stack to have a network layer robust enough to transparently route packets among the various host LAN interfaces which connect to the parallel Ethernet segments.

MANCHESTER ENCODING

Ethernet data is a Manchester encoded 10 mbs signal baseband switched at 10 mHz. See Figure 2. Manchester en- coding begins with a 10 mHz clock. Data are aligned and ex- clusively ORed with the clock. Thus, if the data input line is high on the clock half-cycle, the Manchester encoded out- put line is also switched to high. If the data input line is low on the clock half-cycle, the Manchester encoded output line is also switched to low. The key term is "switched" since if the data line remains in its old state (high or low) the Manchester encoder must "reset" the output line so it may switch to high, or to low, on the clock half-cycle to indi- cate the bit value. Note that Manchester encoding requires that the starting bit to be high and that the receiver will always receive the data one clock half-cycle later. Manch- ester encoding offers data integrity in two ways: since data will arrive at 5 mHz (ones and zeros alternating) or 10 mHz (all ones or zeros), the receiver will always be synchronized with the transmitter. Also, since changing data is half the frequency of no data, Manchester encoding should not misin- terpret spurious voltage shifts as data. Ethernet frame col- lision can now be seen as simply more than two Manchester en- coded data switches within one 10 mHz clock cycle.

FRAME FORMAT

An Ethernet frame consists of the following seven fields (with their size in bytes): Preamble (8), Destination Address (6), Source Address (6), Type [version one Ethernet] or Frame Length [IEEE 802.3] (2), Data (46-1500), Frame CRC (4), and the Interframe Interval of 9.6 microseconds (12). To syn- chronize the transmitter and receiver, Ethernet employs a preamble of 62 alternating ones and zeros followed by two ones (the first eight bytes of each frame). The six-byte ad- dresses yield a maximum address of 2**47-1 or 140,737,488,355,327 possibilities. Addresses are not 2**48-1 because the most significant address bit (if set) indicates that the remaining address is a "multicast." In other words, frames with negative addresses are accepted by a group of stations. If the address is all ones (a negative one), all stations must receive the broadcast frame. In version one Ethernet, the type field was left undefined; but some proto- cols (such as DoD ARPA IP) use the field to indicate the type of frame being sent. Later, IEEE 802.3 redefined the field to be the length of the frame. Therefore, the two versions of Ethernet are electrically compatible, but old style soft- ware may not be compatible if it uses the type field which is subsequently overwritten with the frame length by the Ether- net controller. To minimize this problem, most controllers require the Ethernet driver to place the frame length in the frame length field. The interframe interval (9.6 microsec- onds) is a worst case propagation delay between the two most distant stations on either side of a five segment Ethernet. Thus, by waiting 9.6 microseconds after the cable is idle, a transmitter can assume no other station is using the Ether- net.

Both version one Ethernet and the IEEE 802.3 standards claim the 9.6 microsecond delay is a "rest" interval for the receiving station so that it may have time to prepare for the next frame. The 9.6 microsecond rest interval seems odd be- cause it would constrain the Ethernet protocol to the speed of existing technology. Another possible reason for the 9.6 microsecond rest interval might be to allow for the propaga- tion delay between the two most distant stations and thereby ensure that the cable is really idle. But if one assumes that an electrical signal travels at 200m per microsecond, then the cable-idle argument is flawed. Two hundred meters per microsecond divided into (500m * 5 segments) = 12.5 mi- croseconds (not counting the four repeater/bridge delays). Thus, there appears to be no justification for the interframe delay interval.

Assuming a minimum data frame of 46 bytes, 38 bytes or 45% of overhead is required to send the data. On the other hand, if the data frame is 1500 bytes (the maximum), overhead is 38 bytes or 0.02%.

CSMA/CD PROTOCOL

Ethernet is a carrier sense, multiple access, with col- lision detection (CSMA/CD) data link protocol. The Ethernet transceiver assists in this protocol by providing three sig- nals, transmit, receive, and collision presence to the Ether- net interface. More precisely, collision presence is defined as the detection of two low-to-high or high-to-low transi- tions within the previous 1.6 clock cycles. The Ethernet in- terface logically ANDs the transceiver's collision presence signal with its transmit signal to indicate collision detec- tion. The Ethernet interface also logically ORs collision presence with receive creating a "carrier sense" signal which prevents the station from using the Ethernet while a colli- sion is in progress.

Once carrier sense is asserted, the Ethernet interface monitors the receive line for two successive ones. After the two ones are observed, the frame is assembled in modulo eight until the carrier sense line is unasserted. Note that the IEEE 802.3 Frame Length field is redundant since end-of-frame is determined when voltage switches no longer occur within a 10 mHz clock cycle. If the frame size is not modulo eight, or the CRCs do not match, an error is sent up to the network layer (not back down to the transceiver).

The Ethernet interface will not transmit until 9.6 mi- croseconds after the carrier sense line is unasserted. If two or more stations were waiting for carrier sense to become unasserted, then a collision detection would result and each station must continue to transmit 32 to 48 "jam" bits and wait a random interval before testing carrier sense again. Note that with 32 jam bits, all receivers will hear start-of- jam before the message ends (25.6 microseconds) and thus pre- vent continuous jamming by distant stations which would have not heard the original jam signal before sending their frames. If the Ethernet interface cannot transmit after 16 attempts, and error is reported up to the network layer.

The interval a transmitter must wait after each colli- sion is based on a multiple of 512 bit times or 51.2 mi- croseconds. The interval ranges from zero to 2**k where k is the number of retransmission attempts, or 10, which ever is less. Table 1 shows the mean cumulative delays incurred upon repeated collisions.

Retrans-    Random                                Cumulative
 mission    Number                   Duration      Duration
Attempts    Range     Mean   Bits   (microsec)    (microsec)
--------   --------   ----  ------  ----------   ------------
       1   0 -    2     1      512     51.2          51.2
       2   0 -    4     2     1024    102.4         153.6
       3   0 -    8     4     2048    204.8         358.4
       4   0 -   16     8     4069    406.9         765.3
       5   0 -   32    16     8192    819.2        1584.5 (1.5 ms)
       6   0 -   64    32    16384   1638.4        3222.9
       7   0 -  128    64    32768   3276.8        6499.7
       8   0 -  256   128    65536   6553.6       13053.3
       9   0 -  512   256   131072  13107.2       26160.5
      10   0 - 1024   512   262144  26214.4       52374.9
      11   0 - 1024   512   262144  26214.4       78589.3
      12   0 - 1024   512   262144  26214.4      104803.7 (0.1 sec)
      13   0 - 1024   512   262144  26214.4      131018.1
      14   0 - 1024   512   262144  26214.4      157232.5
      15   0 - 1024   512   262144  26214.4      183446.9
      16   0 - 1024   512   262144  26214.4      209661.3 (0.2 sec)

 Table 1.  Ethernet Cumulative Transmit Delay as a Function of
                     Successive Collisions.

ETHERNET INTEGRATED CIRCUITS

In 1984 four Ethernet chip set designs were announced. Today, two of the four are used in most Ethernet interfaces. Some computer interfaces hide these chips from the programmer with an on-board CPU and OS which exchange messages with the host computer. One of the two sets is the Local Area Network Controller for Ethernet (LANCE) developed jointly by American Micro Devices, Digital Equipment Corporation, and Mostek Cor- poration. The chip set consists of a 48-pin NMOS controller (the AM7990 or MK68590) and a 24-pin serial interface adapter (the AM7991 or MK3891) which performs the Manchester encoding and decoding.

The second chip set was designed by Intel Corporation. Like the LANCE, Intel has two ICs: an Ethernet controller (the 82586) and a parallel-to-serial converter and Manchester encoder-decoder (the 8250). Intel's 82586 Ethernet con- troller offers additional functions over the LANCE. The chip employs a 16-bit internal and two 16-bit external system bus- es. It also has two 8-bit parallel paths to the 8250. One 16-bit external system bus offers four multiplexed DMA paths. The other 16-bit external bus is for receiving commands and transmitting status replys. In 1991, Intel introduced a sec- ond version with 32-bit buses (the 82596DX).

Example controller commands are start, abort, suspend, and resume. Actually, pointers to command lists are passed to the controller which uses them to execute its data link protocol. Example status replys include: idle, active, no receive resources available, command completed, frame re- ceived, and CRC errors. Ethernet frames are passed to the host via a linked list of "buffer descriptors." Additional features of the 82586 over the LANCE are in the area of diag- nostic reports. The chip gathers statistics on its perfor- mance, CRC errors, alignment and lost frames problems, per- cent cable busy, collisions experienced per frame transmit- ted, and the frequency of lost frames due to collisions. The chip also employs time-domain reflectometry to sense open or short circuits and their distance from the transceiver. Fi- nally, the chip has a loop-back feature which allows the host to test the station off-line and a dump command which will cause the 82586 to copy its internal registers into host mem- ory.