Local Area Networks

Early in the microcomputer era, there were a variety of contenders in the LAN arena. However, the dust has now settled, and there are only two important ones left—Ethernet and Token Ring. Others, such as ARCnet and Appletalk, have shares of the market that are vanishingly small.

Ethernet

Ethernet, originally developed at Xerox's Palo Alto Research Center (PARC), is a bus-based local area network strategy. Ethernet is also baseband—only one connection can operate on a wire at one time. The wire is shared through an approach known as Carrier Sense Multiple Access/Collision Detection (CSMA/CD). Ethernet nodes listen to the wire to which they are attached to see if anyone else is broadcasting (carrier sense). If all is quiet, and they want to use the wire, they send out a frame (Ethernet's version of a packet) and then listen again. If another node has also tried to broadcast during this time, their frames will collide and make a distinctive sound (collision detection). If this happens, both nodes will wait a random amount of time and try again. Believe it or not, this weird system works very well. In case you're skeptical, just observe that there are estimated to be on the order of a hundred million Ethernet "nodes" (connected computers or other devices) in operation around the world.

A couple of technical points about Ethernet. First, note that the CSMA/CD approach necessarily limits the length of a network segment. Practically speaking, there has to be some limit on the time a sender waits before it decides that there has not been a collision. The time is related to how long it would take to hear a collision at the farthest point; ergo, there has to be a farthest point. Standard 10 Mbps Ethernet sets this at about 185 meters; we'll see that faster Ethernet has much shorter segments. On a second point, Ethernet addressing is rather clever. Each Ethernet card is given an address, permanently burned into its circuitry, at the factory. In case you're wondering, factories have to register with an international standards body which assigns them a range of numbers to be used. Since the addresses are 48 bits long, there can be trillions of separate ones, something which hasn't been a problem yet. Simple, but very effective.

One reason for Ethernet's success has been that it is an open standard. The IEEE endorsed it, as IEEE 802.3, in ancient times (before 1985). As originally certified, it was only for thick coaxial cable, but as we discussed in Chapter 11, the specification was later extended to cover thin coax and then unshielded twisted-pair. In its original form, Ethernet was specified for no more than 10 Mbps; real world throughput is on the order of 4 Mbps to 6 Mbps. Ethernet's speed was more than sufficient for most networks in the early days of microcomputers. However, as networks got bigger and the traffic from each node got higher, some changes were in order.

The first, and easiest, way of increasing the capacity of an Ethernet network is to segment it. In this approach, one LAN is broken down into two or more smaller ones. Let's say you have 60 nodes and one server. In an unsegmented LAN, all 60 nodes contend with each other for access to the network; collisions increase and the effective speed goes down. If you segment the net, with two LANs connected by bridges, you will greatly increase effective bandwidth. One node on each segment can operate at the same time, which means that you have doubled bandwidth. If a node on LAN 1 wants to communicate with one on LAN 2, its frame is sent through the bridge and only then is the speed of the entire network reduced. Needless to say, in segmenting LANs it is important to balance traffic as much as possible. This raises an important question. Since most traffic will be going not from one node to another, but from a node to a server, where do you put the server?

A Server as Bridge

One approach would be to forget the bridges and put the server on both LANs at once by giving it two network cards (see Figure 12.1). Since the server is now directly on both LANs, it does the bridging. Most servers will support two or three network interface cards; they can even link disparate networks (e.g., Ethernet and Token Ring). The problem here is that the server is now taking on a big load; its CPU and bus will have a lot of network chores to handle in addition to the normal file, print, and application duties.

Multiple Servers

Another option would be to give each LAN its own server. This would not only balance the load, it would help deal with problems of server capacity and the single point of failure issue that obtains in option one. And buying more servers wouldn't be as expensive as it sounds, given that prices are falling. The big problem is management. If you have multiple servers, you will have to deal with the problem of data replication. If users on LAN 1 and LAN 2 are both accessing the same database, there will have to be some mechanism for keeping the copy on Server 1 consistent with that on Server 2. This is a very tricky issue that we'll discuss in greater depth in Chapter 13.

Ethernet Switches

A third approach, the one that is now dominant, is to have both LANs and one or more servers separately connected to an Ethernet switch (Figure 12.2). The switch can give any node a full 10 Mbps pipe to any other node or server. The capacity of the switch matters in this circumstance; if it has a throughput of 100 Mbps, it can handle ten 10 Mbps connections at once. Ethernet switches have all but eliminated hubs in new networks and are the most popular way to speed up legacy LANs. Configuration is important, though. If, as happens with most LANs, nearly all traffic is from individual nodes to a server or servers, it won't help to have a single 10 Mbps connection to the server. You need either multiple 10 Mbps links (good), or one or more 100 Mbps connections (better), or a Gbps one (best, but not widely available yet). Also note that, as shown in Figure 12.2, the LAN's switch can also connect to other LANs through an additional switch or router.

Two points in passing. For a time, networks of the type described in option three used routers rather than switches. Switches won out in part because they are faster, cheaper, and simpler to manage and in part because connecting local networks doesn't require the flexibility in making choices that a router provides.

Fast(er) Ethernet

Our final strategy for a faster LAN is to have the links operate at rates greater than Ethernet's original 10 Mbps. Once it became clear that the higher speed was needed, multiple players vied for the right to be the leader in developing a new standard. A lot of approaches appeared, all technically feasible (as Robert X. Cringley has observed, for a market in the millions, there are people in Silicon Valley who could make Ethernet work over barbed wire). The eventual winner goes by the creative name of Fast Ethernet.

Also called 100Base-T, Fast Ethernet uses either two (typical) or four pairs of Category 5 UTP; some specifications allow Category 3 (four pairs only), but this is rare. Cable runs for 100Base-T are limited to about 200 meters. Because Ethernet is not only very popular but well understood, vendors moved quickly to develop hardware to support it. It took only a few years for the price of Fast Ethernet cards to fall to that of 10Base-T. Indeed, a very common network card now is one that supports both standards at once. The price differential in hubs and switches is greater, but these costs are coming down as well. The 200 meter segment limit, which might be a problem for many LANs, is offset by the fact that newer LANs use switches rather than hubs; this means that the distance limitation for a segment applies only to the cable from a node to a switch rather than the distance between the two farthest points of the entire LAN. It seems clear that Fast Ethernet will replace 10Base-T as the most common LAN network system.

Ethernet without Wires The IEEE has recently created standards for versions of Ethernet that use wireless connections. The most important of these, IEEE 802.11b, appears to be headed for widespread use. The new 802.11 standard has several advantages: 1) it uses a similar technology to wired Ethernet, which will make it easy and inexpensive to link to existing Ethernet networks; 2) it uses an unlicensed part of the spectrum, the 2.4 GHz band, that simplifies setup, and it employs the CDMA/spread spectrum transmission method that is highly resistant to interference. Apple, which calls the technology AirPort, started the market moving with products that were substantially less expensive than competing technologies.

An alternative strategy to 100Base-T is called 100VG AnyLAN. It is not real Ethernet because it doesn't use Ethernet's CSMA/CD access mode. Instead, AnyLAN uses an intelligent hub-based system called demand priority that does a better job of providing quality of service for streaming data such as video and audio; it also integrates well with both 10Base-T and Token Ring legacy LANs. Despite support from Hewlett-Packard, AnyLAN has not found much acceptance in the market—Fast Ethernet is the logical next step for the overwhelming majority of users.

Token Ring

After Ethernet, the only other major LAN network scheme is Token Ring. Developed by IBM, Token Ring uses the token passing scheme described in Chapter 11. Token ring LANs don't use the dual counter-rotating ring strategy popular with wide area networks like FDDI and SONET (see the last section of this chapter). Instead, a Token Ring LAN uses a star topology. The hub has the responsibility for maintaining the integrity of the ring; if a node or a link to a node fails, the hub drops it out immediately and restores the ring.

Token Ring was introduced by IBM at 4 Mbps in the early 1980s. Like Ethernet, it was made an IEEE standard (802.5). IBM's preference has been for shielded twisted-pair cable, which, in its two-pair incarnation, is called Type 1. Other cabling systems, including UTP Category 3, can be used but are not common. Later, Token Ring was upgraded to 16 Mbps (Type 1 cable only). The nominal speed of these LANs (actually of any network system) can be deceptive, though. Ethernet at 10 Mbps normally operates at about 4 Mbps to 6 Mbps, while Token Ring's throughput is normally closer to its actual rating. Generally speaking, Ethernet is more efficient with bursty traffic (irregular, often long transmissions) while Token Ring does better with continuous small transactions.

IBM now provides 100 Mbps Token Ring gear. Together with the appearance of Token Ring switches, this will allow the technology to retain most of its market share. There is even talk of Gigabit Token Ring.

Differences in speed haven't been the reason why Ethernet has those hundred million or so nodes installed worldwide while Token Ring has a fraction of that. The reason is economics. Although Token Ring is an international standard, IBM holds patents on most of the technology. And while IBM said that it would license Token Ring broadly, it was slow to do so. The result was that Token Ring hardware was available from relatively few vendors and therefore at prices much higher than Ethernet. Token Ring is mostly used at IBM shops because IBM supports it, not only on microcomputers, but in the entire range of its advanced mini and mainframe systems. There's nothing wrong with Token Ring; it's just that quasi-proprietary technologies usually take, at best, a market niche. Don't weep for IBM, though. Estimates are that there are more than 10 million Token Ring nodes out there.