HighSpeed TCP for Large Congestion Windows by Sally Floyd Notes

• Modifies TCP congestion control mechanism for use with large congestion windows
• TCP constrains congestion windows that can be achieved in realistic environments
  
• Standard TCP average congestion window is
  1.2/sqrt(p) segments
  where p is loss.
  
• 83000 segments = 1 Gb
  
• Do not ramp up more quickly than allowed by slow start. It would necessitate more explicit feedback from routers along path.
  
• Simpilify to packet loss rates of at most 10^-2
  
• In standard tcp: need w/1.5 rtts between loss events.
  
• Implement:
  Low_Window: threshold of which response function to use - highspeed or standard
  High_Window: average congestion window to be achieved.
  High_P: the tcp packet drop rate.
  
• Need that response function gives straight line on log-log scale. Gives response curve of:
  W = (p/Low_P)^S Low_Window
  Where Low_P corresponds to the packet drop rate of Low_Window.
  S = (log High_Window - log Low_Window) / (log High_P - log Low_P )
  Eg for Low_P = 10^-3, Low_Window = 38 and High_Window = 83000, High_P = 10^-7, W = 0.12 / p ^0.835
  
• Appropiate response function is one that gives a plausible packet drop rate for a connection throughput of 10gb/s.
• Fairness: Considered through ratio of window sizes for different p
• Need AIMD parameters of a(w) and b(w)
  
• Response to ACK: w -> w + a(w)/w
  
• Congestion: w -> ( 1 - b(w) ) w
  
• Standard TCP: a(w) = 1, b(w) = 0.5 for all values of w.
• When w = High_Window, we have loss rate High_P. so,
• a(w) = High_Window^2 * High_P * 2 * b(w)/( 2 - b(w) )
• High_Decrease: decrease parameter for b(w) when w = High_Window.
  Eg. High_Decrease = 0.1, High_Window = 83000, High_P = 10^-7 => b(83000) = 0.1, ie a decrease of 10% after a congestion event. Also a(83000) = 72, if increase of 72 segments or under 0.1% of W.
• Need b(w) when w > Low_Window:
• b(w) = ( High_Decrease - 0.5 ) ( log(w) - log(W) / ( log( W_1) - log(W) ) ) + 0.5
• Giving
  a(w) = w^2 * p(w) * 2 * b(w) / ( 2 - b(w) )

Thu, 7 August, 2003 3:25