TCP Congestion Control

• Congestion occurs when total arrival rate from all packet flows exceeds R over a sustained(维持) period of time
• Buffers(缓冲) at multiplexer will fill and packets will be lost

Phases of Congestion Behavior

1. Light traffic
   • Arrival Rate << R
   • Low delay
   • Can accommodate(容纳) more
2. Knee (congestion onset)
   • Arrival rate approaches R
   • Delay increases rapidly
   • Throughput(吞吐量) begins to saturate(饱和)
3. Congestion collapse
   • Arrival rate > R
   • Large delays, packet loss
   • Useful application throughput drops

Congestion Window

• Desired operating point: just before knee
• TCP sender maintains a congestion window (cwnd) to control congestion at intermediate(中间的) routers
• Effective window is minimum of congestion window and advertised window(广播窗口)
• Problem: senders does not know what its “fair” share of available bandwidth should be
• Solution:
  • adapt dynamically to available BW
  • Senders probe(探查) the network by increasing cwnd
  • When congestion detected, senders reduce rate
  • Ideally, sending rate stabilizes(稳定) near optimal(最优) point

Congestion Window (Cont.)

• How does the TCP congestion algorithm change congestion window dynamically according to the most up-to-date state of the network?
• At light traffic: each segment is ACKed quickly
  • Increase cwnd aggresively
• At knee: segment ACKs arrive, but more slowly
  • Slow down increase in cwnd
• At congestion: segments encounter large delays, timeout, segments are dropped in router buffers
  • Reduce transmission rate, then probe again

TCP Congestion Control (1): Slow Start

• Slow start: increase congestion window size by one segment upon receiving an ACK from receiver
  • initialized at 2 segments; usually 1 segment
  • used at start of data transfer
  • congestion window increases exponentially(指数)

TCP Congestion Control (2): Congestion Avoidance

• Algorithm progressively(逐步) sets a congestion threshold(门槛)
  • When cwnd > threshold, slow down rate at which cwnd is increased
• Increase congestion window size by one segment per round-trip-time (RTT)
  • Each time an** ACK arrives, cwnd is increased by 1/cwnd**
  • In one RTT, all ccwnd segments are sent, so total increase in cwnd is cwnd x 1/cwnd = 1
  • cwnd grows linearly with time

TCP Congestion Control (3): Congestion

• Congestion is detected upon timeout or receipt of duplicate ACKs
• Assume current cwnd corresponds to available bandwidth
• Adjust congestion threshold = ½ x current cwnd
• Reset cwnd to 1
• Go back to slow-start
• Over several cycles expect to converge(收敛) to congestion threshold equal to about ½ the available bandwidth

Fast Retransmit & Fast Recovery

• Congestion causes many segments to be dropped
• Burt if only a single segment is dropped, then subsequent(随后的) segments trigger duplicate ACKs before timeout
• Can avoid large decrease in cwnd as follows:
  • When three duplicate ACKs arrive before timeout expires(期满), retransmit lost segment immediately
  • Reset congestion threshold to ½ cwnd
  • Reset cwnd to congestion threshold + 3 to account for the three segments that triggered duplicate ACKs
  • Remain in congestion avoidance phase
  • In absence of timeouts, cwnd will oscillate(振动) around optimal value

TCP Congestion Control: Fast Retransmit & Fast Recovery