A component may be operating sub-optimally at the physical layer because it
is being utilized at a higher average than it is configured to operate. When
troubleshooting this type of problem, it will become evident that resources for
the device are operating at or near the maximum capacity and there is an
increase in the number of interface errors. Gathering symptoms reveals
excessive runts, late collisions, or an increase in the number of buffer
failures. The output from a ping or
traceroute command results in excessive packet loss or
latency.
How Much Utilization is ok?
Shared Ethernet networks are believed
to suffer from throughput problems when average traffic loads approach a
maximum average capacity level of 40 percent. This percentage is actually
conservative. Higher average percentages are certainly possible. The classic
solution to excessive traffic is to micro-segment LANs by installing switches.
This solution works well until the amount of broadcast traffic grows too large.
Since bridges and switches always forward all broadcast traffic to all ports,
even infrequent broadcasts from each station will eventually be too many when
the station count for the broadcast domain goes up.
If a single station
is connected to a half duplex switch port acceptable utilization is best
learned by monitoring switch port statistics. There will be some level of
collisions, but since there are only two devices on that link (the switch and
the station) the utilization level should be capable of averaging quite high.
There will likely be excessive collisions errors reported by the switch
periodically, resulting from the Ethernet capture effect. This problem is not
terribly significant overall, and will result in a slight reduction in
performance. If the link is allowed to negotiate to full duplex instead of
half, the connection should be capable of approaching theoretical limits for
full line rate Ethernet. This will depend on the processing power of the
attached station. The current generation of switches is quite capable of
sustaining line-rate traffic at the minimum frame size.
Another problem
can occur when access to servers or services is reached through a single switch
uplink path. Unless the bandwidth of the uplink path is bigger than the total
of simultaneous station requests, the uplink itself becomes a bottleneck. This
scenario arises when a network is designed with all servers collected in a
server farm, separated from the VLANs they serve.
Network bottlenecks or
congestion typically manifests itself to users with the following symptoms:

- Highly variable response times
- Network time-outs or server disconnects
- Inability to establish network connections
- Slower application loading and/or running
Even when network resources are on the same LAN segment as the users,
it is still necessary to examine the network architecture to see if bottlenecks
exist. If too much traffic is required to pass through an inadequate
aggregation path then a potentially useful architecture design is defeated. To
prevent this, the network administrator should limit inter-segment traffic by
carefully considering which nodes should attach to each segment. The process
involves an investment of time and may need to be repeated regularly on
extremely dynamic networks. Monitor uplink paths as a part of routine
maintenance in order to detect impending saturation of any one path.