In CSMA/CD, it is important to ensure that the collision window is not a large fraction of overall packet size. When transmission speeds increase, this can only be achieved by either increasing average packet size or restricting maximum propagation delay (by constraining network size). The former option is undesirable since using different packet lengths would compromise compatibility with standard Ethernet so it is judged preferable to restrict the network size, by limiting cable runs to 450m (optical only) maximum. 802.3u specifies only a hub implementation of fast Ethernet, 100Base-T, and there is no equivalent of 10Base5 or 10Base2. Restricting maximum propagation delay also allows the minimum packet size to remain unchanged. Recall that the minimum packet size in standard Ethernet is chosen so that any collision will always be detected before transmission is complete. If the transmission rate is a factor of ten faster, the cross-network propagation delay must be reduced by a factor of ten to compensate. Note, that these restrictions do not apply to switched full-duplex Ethernet, where collisions do not occur.
In 100Base-T, stations are attached to a central hub or switch using twisted pair or fibre optic point-to-point links. The commonest implementation is 100Base-TX which uses two high-grade Category 5 UTP cables, one for each direction Full duplex operation requires the use of a 100Base-T switch, as opposed to a simple hub, which merely mimics the operation of a single shared bus. Other implementations which are much less common are 100Base-T4 using 4 pairs of Category 3 UTP for half-duplex operation only and 100Base-FX, which uses two multimode optic fibres, operating like 100Base-TX, but allowing stations to be up to 450m away from the hub in half-duplex mode. In full-duplex installations, 100Base-FX allows a station to be up to 2km from its switch. Another variant, not part of the original 802.3u, is 100Base-T2 which uses sophisticated equalisation and noise cancellation techniques to support multilevel signalling on two Category 3 twisted pairs.
The general operation of Fast Ethernet is similar to that of its slower analogue, although it operates ten times faster: for example, at 10Mbps, frames must be separated by an inter-frame gapof 9.6ms; in 802.3u, this is reduced to 960ns. The whole point of 100Base-T is to make migration from 10Base-T as easy as possible, so most 100Base-T hubs and switches can handle a mix of stations operating at both 10 and 100Mbps. Likewise, most 100Base-T station adapters (Ethernet cards) can be connected to a 10Base-T or 100Base-T port, and will adjust themselves to the correct speed automatically.
In 802.3u the separation into MAC and physical layers is similar to 10Mbps 802.3 but the physical layer structure has been redesigned. At 100Mbps, Ethernet uses much more complex signal encoding mechanisms than simple Manchester code and these differ from one variant to another. As a result the PMA sublayer is replaced by a more complex, medium-dependent mini-stack known as the PHY. The PHY has three sublayers, the physical coding sublayer (PCS), physical medium attachment sublayer (PMA) and the physical medium dependent sublayer (PMD), which perform the signal encoding and decoding carried out in 802.3 by the PLS as well as the lower level medium-dependent functions. Because of the variety of options available in 802.3u, including a 10Mbps backward-compatibility mode, encoding and decoding are medium dependent, so the new division of responsibilities is actually quite logical.
Different 100Base-T media (FX, TX, T4, T2) must implement different PHYs but each is capable of 10Mbps 802.3 compatible communication. An 802.3u station uses a process called autonegotiation, also a PHY responsibility, to decide which means of communications should be used (10 or 100Mbps; full or half duplex). At the top, the PHY connects to the medium-independent portion of the physical layer via an interface called the medium independent interface (MII) which is much more complex than the AUI and involves the transfer of 4-bit parallel data (at 25MHz in 100Mbps operation) as well as control signals; at the bottom it connects to the physical medium via a medium dependent interface (MDI). Above the MII is a reconciliation sublayer (RS) which is medium independent and responsible for reconciling the MII with the 802.3u MAC sublayer, so that the latter sees the same physical layer service as 802.3.
Figure 1: Comparative architectures of 802.3 and 802.3u
In July 1997 IEEE issued a draft standard (IEEE 802.3z) defining 1000Mbps Gigabit Ethernet also known as 1000Base-X. Gigabit Ethernet may be hub or switch-based although almost all installations use switches. Its architecture is almost identical to 802.3u although the physical layer is modified slightly: for example the MII is replaced by a Gigabit MII (GMII) which is 8-bit parallel and runs at 125MHz. The Gigabit PHY sublayers support several types of media: 850nm laser over multi-mode fibre in 1000Base-SX, over a distance of up to 550m; 1300nm laser over single-mode or multi-mode fibre in 1000Base-LX up to 3km; 150 ohm STP (shielded twisted pair) in 1000Base-CX up to 25m full-duplex; and (as defined later in IEEE 802.3ab) 4 pairs of Category 5 UTP up to 100m, in 1000Base-T. 1000Base-T uses 5 level signalling at 125MHz to achieve the 1000Mbps transfer rate (half-duplex) but neither it nor 1000Base-CX are widely used.
The MAC layer in Gigabit Ethernet is just as in standard and fast Ethernet with some minor modifications. The most important of these relates to solving the collision window problem which if tackled in the same manner as in 802.3u would result in networks too restricted in size to be of use. Although the logical minimum frame length of 64 bytes is maintained, actual physical frames are always padded out to a real minimum of 512 bytes to allow longer cable distances to be supported on non-switch (hub) implementations. This is called carrier extension. However, if a large number of small frames were sent, this would result in a large overhead greatly reducing performance. To avoid this, a technique known as packet bursting has been introduced. This allows a source with a number of small packets to send, to concatenate them (with a small inter-packet gap) up to a maximum length of 1500 bytes. It should be said however that the collision window problem is only an issue in half-duplex Gigabit installations and since these are not common, the carrier extension and frame bursting techniques are not usually required.
As with 100Base-T, the great advantage of 1000Base-X is that, as an extended implementation of 802.3, it can easily coexist with earlier forms of Ethernet.