Figure 4: Avoiding Congestion
The problem of congestion always exists when a progressive bandwidth sharing strategy is used. The effects and handling of congestion in switches differ depending on the equipment used (compare e.g. ). We assume that congestion is problematic for the traffic in the network in any case and should therefore be avoided where possible.
Congestion is indicated by some increase in the used bandwidth on the output links and depending on the used equipment on certain buffer values. For generality, we only consider the amount of used bandwidth so far. We have to determine a limit, which surely points on an approaching congestion. This is quite complex, as the increase in the use of bandwidth could be very sudden and would make it impossible to act in time.
As previously, a possible action value is related to usually active connections and their requested resources. If there are currently less connections active than usual and more connections are configured than active (i.e. we have idle connections) the probability of congestion is very high if the resources left on a link are less than the mean resources which the idle connections request. However, this is only one possible indication of congestion. Further approaches, which include vendor specific management parameters as well, are topic for further research.
Again, the action to perform is the renegotiation of existing connections according to the already introduced fairness policies. If no renegotiation is possible and congestion appears often the physical redesign of the network structure should be considered.
After such an occurrence, we should remember this situation and use the knowledge to manipulate the admission control accordingly (second branch of Figure 4). That means a report has to be generated, where the time and traffic configuration which led to the (potential) congestion is registered. The CAC then has to be adapted in order not to accept this traffic mix (possibly in conjunction with the sources who send the traffic) at this time again.
The control part in this case will be a negative message, i.e. there was a congestion on a particular link in spite of the renegotiation of connections. Note that the detection of a potential congestion can be seen as the control part for the second branch in Figure 4.