Upgrading and Repairing Networks (5th Edition)

Because traditional Ethernet uses a shared network medium, tracking down problems can at times be difficult. Problems can arise from simple things such as bent or broken cables, to loose connectors, and faulty network adapters. The most common problem you will find, however, is in controlling collision rates as the network size increases.

Collision Rates

Keeping the network healthy requires that you be sure that all physical components are functioning normally and at optimal levels of performance. However, you still need to monitor the network to be sure that other factors are not limiting the amount of real data that can travel through the network.

Although collisions are a normal event in legacy Ethernet networks, and indeed are expected for a network based on Ethernet technology, it is always possible for excessive collisions to cause a significant degradation of performance that will be noticeable to end users.

Collisions and Network Utilization

When a device begins to experience collisions at a rate that is 1% of the total network traffic, you might have a problem. Another statistic to watch when monitoring the network is utilization. In theory, you might expect that a network operating at 10 million bits per second would actually be capable of transmitting that much data on a continuous basis. However, that is not the case. In most Ethernet networks, the actual utilization rate is only 40% before performance begins to degrade rapidly. As utilization rises, so do collisions.

Note

Keep in mind that in this section collisions are discussed, and collisions occur only on shared network media. If switches are used (in full-duplex mode), then collisions do not occur. If you are using hubs, it might be a simple solution to increase your network bandwidth utilization by simply replacing older hubs with newer switches.

If the network topology rules are followed and the network utilization is low, excessive collisions might be due to a faulty network card that is not listening to the network. You can find more information on this scenario later in this chapter in the section "Faulty Network Adapters."

Detecting Collisions

A simple method for determining how many collisions are occurring is to look at the LED lights on the hub or switch. Most hubs have an LED that lights up when a collision is detected. If you notice that this light is flashing continuously or very frequently, investigate further to determine whether the rate is excessive. If it is, take action to reduce it. Using network monitoring software, you can determine the utilization rate of the network. When you get above a 30%40% utilization rate, it's time to start thinking of segmenting your LAN into smaller collision domains.

LAN analyzers and monitoring tools can aid you in counting the number of collisions that are occurring. Management consoles that employ SNMP and RMON probes can be useful for collecting statistical information used to localize segments in the network that experience high collision rates. The historical data maintained by RMON can be analyzed and stored for use in creating baseline data that you can use to judge network performance. If you are about to purchase a new switch or hub, check the documentation to see whether it supports telnet management sessions. This is a common feature on even low-end hubs now. In a small network with only a few switches, using built-in management software is a lot cheaper than investing in network management software like SMS or HP OpenView.

Collision Types

A good network analyzer gives you a lot of statistical information. When it comes to collisions, there will most likely be more than one kind of statistic to help point out the cause of the collision.

Local Collisions

A local collision (also called an early collision) is a collision that happens on the local segment during the transmission of the first 64 bytes of a frame. This is the most common type of collision you will see on a network segment, and usually does not indicate a hardware problem. This type of collision happens when two different stations on the LAN detect that nothing is being transmitted on the wire and both begin to transmit at about the same time. The result is a frame called a runt, named because only part of the frame was transmitted successfully before the collision event occurred. The Ethernet specifications take into consideration this expected event, and both stations use the backoff algorithm to delay transmission.

When high levels of early collisions are occurring, look to see whether the utilization on the segment is nearing or surpassing 40%. If this is the case on a regular basis, the segment is probably overloaded. Consider using a switch to limit collisions. If you can identify a particular node that is experiencing a high rate of local collisions, there might be a hardware problem. Check the connectors that join it to the network; if no fault is found there, try replacing the network adapter card to see whether that is the problem.

Late Collisions

A late collision occurs when two devices on the network start to transmit at the same time and do not detect the collision immediately. A network segment that is too long usually causes this kind of collision. If the time it takes to put the frame on the network is shorter than the amount of time it takes for the frame to travel to the node that is the greatest distance away, neither device will know that the other has started transmitting until after the first 64 bytes (the minimum frame size) have been put on the wire.

For example, suppose that workstation A begins to transmit a frame and finishes transmitting before the signal reaches workstation B, which has been cabled to the network at a distance that exceeds the specs. Workstation B, thinking that the wire is clear, begins to transmit its frame just before the signal from workstation A reaches it. Of course, because workstation B is closest to the collision event, it detects the collision. However, because workstation A has finished transmitting the frame, it has also stopped listening to detect whether a collision has occurred. The end result is that workstation A thinks it was able to successfully transmit the frame. It has no idea that a collision has occurred.

Late collisions do not cause a frame to be retransmitted, simply because the NIC does not know that a collision has occurred. It is up to a higher-level protocol to determine that something has caused an error and to request retransmission.

If the LAN is experiencing high levels of late collision events, check for topology problems. This includes not only excessive cable lengths, but also using too many repeaters or other devices. If no apparent problems are found and the network appears up to specifications, there is probably a hardware problem. Try to locate the offending NIC or cable by looking at the addressing information decoded by a LAN analyzer.

Sampling Intervals

When monitoring for collisions, don't jump to conclusions when you see only sporadic increases. Take samples several times during the workday, and try to correlate them with the functions being performed by users on the network at that time. Sometimes it is the actual day that matters, and not the time. For example, at the end of a month or a quarter, many business functions are performedsuch as accounting reportsthat generate large amounts of network use. It is a simple matter to determine which month-end tasks need to be done first and in what order. Sometimes scheduling is all that is necessary to solve a network congestion problem.

An overall average of the number of collisions that occur per second, along with the network utilization rate, is useful in determining whether the network is becoming saturated. Information about peak levels is useful for designing user work patterns so that the network is used more efficiently.

Reducing Collisions

There are several reasons why collisions will occur at excessive rates. Some of those reasons include ignoring topology rules, faulty hardware, and an overloaded segment (too many users).

Incorrect Network Topology

If you use segments that exceed the length permitted by your network topology, some devices on the network might not detect that the network is in use until a transmission by another node is well underway. Check your cable lengths and be sure they are within the standards. When it comes time to expand the LAN, you should never haphazardly add new segments by simply attaching a new repeater, hub, or bridge to the network. For this reason, it is important to keep an up-to-date map of the physical topology of the network so that you can plan additions before you implement them.

Remember that, for 10BASE-T, workstations can be no farther than 100 meters from the hub. In addition, the 5-4-3 rule states that there can be a maximum of five cable segments on the LAN, with a maximum of four repeaters or concentrators, and only three segments can have nodes attached. For Fast Ethernet and Gigabit Ethernet, be sure that you don't exceed the topological metrics imposed by the physical network media used.

Faulty Network Adapters

One particular problem is an adapter that does not sense the carrier signal due to faulty hardware, and begins to transmit whenever it wants to, thinking that the wire is available. In Chapter 7, "Network Interface Cards," you will find a more detailed discussion of troubleshooting NICs. However, a basic strategy to follow is to replace the suspect device and, if that does not solve the problem, try to use a different cable to connect the NIC to the network or try to reseat the NIC in another slot in the computer. When replacing the device, be sure to use a substitute that is known to be in good working order. The same goes for replacement cables. Another troubleshooting tactic is to use the diagnostic software provided by the network adapter's manufacturer.

Top Talkers

There are only so many devices you can place on a network in the same broadcast domain before performance begins to suffer. A small number of high-performance computers that generate a lot of network traffic can produce the same result. Remember that as utilization rises, so do collisions. So when you are experiencing a high collision rate and the network segment's utilization approaches or exceeds the 40% mark, it's time to consider segmenting the LAN using a switch or similar device. A switch, which can be used to give high-end servers a full-duplex connection, is an ideal choice when a local segment contains both end users and powerful servers that are "top talkers."