Congestion Control

• Last time, I briefly reviewed the need for congestion control.

  • When routers/switches get overloaded, they drop packets.

  • If the hosts whose packets get dropped simply react by sending duplicates, the congestion is likely to increase (in a vicious cycle).
  and the difference between congestion control and flow control.

• Congestion control is a resource allocation problem. We are allocating the transmission and buffering capacity of network switches to handle the flow of packets from end stations.

• Like many allocation problems, there are two basic approaches:

  • Avoidance/Prevention -- come up with allocation policies that make congestion impossible

    • This is a reasonable approach in a network based on virtual circuits. In such a network, when one machine makes a connection to another, all intermediate switches can reserve capacity for the connection (or reject the connection if the resources are not available).

  • Detection and Recovery -- include mechanisms that allow the network to detect the onset of congestion and reduce network usage in response.

    • This approach is more appropriate in the TCP/IP world where there are no explicit connections at the lowest level (IP).

• The prevention approach tends to waste resources on average since many connections will not continuously used the capacity allocated to them.

  • To avoid congestion, each router will need to reserve enough capacity for the sum of the maximum bandwidths of all outstanding connections even though most connections will transmit below their maximums most of the time.

• The recovery approach will waste resources during the transient periods where the network does become congested.

• Given our loyalty to TCP/IP, we will consider the detection/recovery mechanisms built into TCP for congestion control.