Congestion

• When the number of packets traveling through a network approach or exceed the handling ability, congestion occurs.
• The study of congestion control is based on queuing theory.
• Each node on the network contains a queue.
  • Arriving packets are stored in an input buffer
  • Eventually the processor observes these packets
    • To determine the destination.
    • It must parse this from the packet header information
    • It will place the packet in the proper output queue.
  • Packets are then transmitted from the output queue.
• Queues are finite.
• If the incoming rate exceeds the outgoing rate there is definitely a problem.
• As the arrival rate approaches the transmission rate, problems begin to develop.
• We are concerned about queue delay and throughput
• When the queues become full there are two choices:
  • Drop excess packets
  • Tell the source to slow down.
• Figure 13.3 assumes infinite buffers.
• 
• Note, as we approach 100% load, the delay goes up exponentially
• This is because packets must wait longer and longer in the output queue.
• But we don't have infinite buffers.
• In an actual network other factors come into play
  • As delay increases, retransmission becomes a factor as packets and acks are lost
  • Messages about delays begin to hit the network and route costs change
  • Attempts to route through other sources increases time spent in routing algorithms.
  • ICMP Source Quench (and others) messages begin flowing adding to traffic
  • Other nodes begin experiencing queues filling as well.
• Figure 13.4 has a real-world look at congestion.
• 
• Note after point B in this figure, even delivered packets are assumed to be lost. The application has most likely timed out.