When the load on a network is so high that it results in large numbers of collisions and lost frames, the productivity of the users is greatly reduced. This is called congestion and it can be solved in one of two ways: either scrap the entire network currently in place and replace it with a faster one; or install an Ethernet switch to create multiple small networks.

In a "traditional" Ethernet network, there is 10 Mbps of bandwidth available. This bandwidth is shared among all of the users of the network who wish to transmit or recieve information at any one time. In a large network, there is a very high probability that several users will make a demand on the network at the same time, and if these demands occur faster than the network can handle them, eventually the network seems to slow to a crawl for all users.

An excellent analogy for this situation is to look at a road between two points. At 4:00 a.m. there is usually very little traffic, and anyone who does travel can get to his or her destination quickly and unimpeded. However, at 8:00 a.m., rush hour hits and suddenly there are more cars on the road that can be handled. The result is that traffic slows to a crawl, and a trip that would take ten minutes at 4:00 a.m. now takes over an hour to complete. The road itself and the cars have not changed, however there are now too many cars using the road at one time to be accomodated properly.

Now, let's look at a redesign of the highway having the congestion problem. If we built a dedicated road from each person's driveway directly to his or her destination. then the rush hour congestion problem would go away. Each person would be able to drive at full speed at all times, and all of the other traffic would be irrelevant. It is impossible to set this up in a highway system, however it is possible to do it in a network.

Switches allow us to create a "dedicated road" between individual users (or small groups of users) and their destination (usually a file server). The way they work is by providing many individual ports, each running at 10 Mbps interconnected through a high speed backplane. Each frame, or piece of information, arriving on any port has a Destination Address field which identifies where it is going to. The switch examines each frame's Destination Address field and forwards it only to the port which is attached to the destination device. It does not send it anywhere else. Several of these conversations can go through the switch at one time, effectively multiplying the network's bandwidth by the number of conversations happening at any particular moment.

Another analogy which is useful for understanding how switches increase the speed of a network is to think in terms of plumbing. For sake of argument, assume that every PC on a network is a sink, and a 10 Mb/s connection is a 1/2-inch pipe. Normally, a 1/2-inch pipe will allow enough water to flow for one or two sinks to have enough water pressure to fill quickly. However, putting more sinks on that same 1/2-inch pipe will drop the water pressure enough that eventually the sinks take a very long time to fill.

To allow all sinks to fill quickly, we can connect the source of water to a larger (6-inch) pipe, and then connect each sink to the 6-inch pipe via its own 1/2-inch pipe. This guarantees that all sinks will have enough water pressure to fill quickly. See Figure One for an image of this concept.

Figure One
"Fat Pipe" Increasing Performance

The solution is to widen the road to our shared resource so that it can support the full load of most or all of the individual roads at once. In other words, we increase the bandwidth to our servers while connecting our clients at 10 Mbps. This is usually referred to as a High Speed Backbone. In networking slang, it is commonly called a "Fat Pipe."

A high speed backbone network is usually run at a speed of 100 Mbps, and is used to interconnect all of the servers and switches on the network. A diagram of such a setup is shown in Figure Two.

Figure Two
Switching With High Speed Backbone

In Figure One we have two File Servers, two "Power Users" who regularly tranfer large files, and ten "Undemanding Users" who mainly use the network for only an occasional print job or e-mail message (due to space constraints, only two of the "Undemanding Users" are actually shown in the drawing).

This layout is splitting our overall network into four subnetworks. From left to right these subnetworks are outlined in Red, Green, Blue, and Violet. The Red subnetwork is a shared 10 Mbps setup, with all of the "Undemanding Users" sharing 10 Mbps of bandwidth. The Green and Blue subnets are dedicated 10 Mbps connections, sometimes referred to as "Private Ethernets." Here, each of the two power users has 10 Mbps of bandwidth dedicated to his or her machine, and this bandwidth is not shared with anyone else. Finally, we have our Violet subnetwork. This one is a Fast Ethernet setup running at a speed of 100 Mbps, and the bandwidth is shared by the two servers.

This is the most common way of setting up a switched network, and almost always results in an optimal price/performance ratio. We limit the amount of expensive Fast Ethernet hardware needed by only using it where its cost is justified by the performance it gives in handling the load at that point in the network, while leveraging an existing investment in 10 Mbps equipment in less demanding parts of the network. As a 10/100 switch is a fairly costly piece of equipment, each port we dedicate to a user is also rather expensive, so again these are only dedicated to individual users where that user's load justifies it. Finally, we can set up shared subnetworks which lump anywhere from two up to 100 users on one switch port.

The key to success in setting up a high speed backbone network such as this is to properly balance the demand of the users to maximize the number of users on each port while still maintaining high performance. It has been found that 12 users on one switch port is the maximum allowing superb performance. However, this number will vary widely depending on each user's usage patterns.

Store and Forward

A Store and Forward switch operates much as its name implies; first it stores each incoming frame in a buffer, checks it for errors, and if the frame is good it then forwards it to its destination port. The advantage of this type of switch is that it prevents wasting bandwidth on the destination network by invalid or damaged frames. The disadvantage is that it increases the latency of the switch slightly. In a network with few errors, this results in lower overall throughput through the switch. Store and Forward is most useful in networks which may experience high error rates.

Cut Through

A Cut Through switch operates differently than a Store and Forward type. In a Cut Through switch, the switch begins forwarding the frame immediately upon recieving the Destination Address. This results in a very low latency and is somewhat faster than a Store and Forward switch, as each frame is in the switch for less time. However, this scheme can propagate errors from one subnetwork to another, which can result in bandwidth being wasted in the forwarding of invalid or damaged frames. Cut Through switches work best in networks which experience few errors.

Hybrid

A Hybrid switch is an attempt to get the best of both Store and Forward switches and Cut Through switches. A Hybrid switch normally operates in Cut Through mode, but constantly monitors the rate at which invalid or damaged frames are forwarded. If these errors occur at a frequency higher than a certain threshold value, the switch then stops operating as a Cut Through switch and begins operating like a Store and Forward unit. If the error rate drops back below the threshold, then the switch will again go into a Cut Through mode. This gives the performance advantage of Cut Through switches when error rates are low and the error trapping of Store and Forward switches when error rates are high.

Analisys of the network will most likely find that you have a large number of users who are not going to place a heavy load on the network, and a smaller number of users who will place a large load on the network. We now group the Undemanding Users together on a hub and connect each hub to a switch port. Our more demanding users will usually be either directly connected to the switch, or if they are on hubs, fewer of them will be sharing each switch port than on the Undemanding User portion.

One part of the design which requires serious thought is the choice of the technology for a High Speed Backbone. If the backbone is expected to carry a heavy load, then the preferred technology to use for it is FDDI, followed closely by 100VG-AnyLAN. Both of these are deterministic in nature, and allow a greater network utilization than networks built with a Fast Ethernet backbone, which is used where utilization is somewhat lower. Please consult our pages on each technology to help you to decide on a high speed backbone type.

One point which should be kept in mind regarding the design of a switched network is that traffic patterns vary by user and time. Therefore, just taking a "snapshot" of network usage patterns may lead to the wrong conclusions and result in a design which is not optimal. It is always advisable to monitor usage patterns over a period of several days to a week to decide how to allocate network bandwidth optimally. Also, in almost all cases, a process of trial and error may be required to fully optimize the design.