FACK TCP is using SACK option for better recovery. It decouples the congestion control from data recovery algorithms, i.e. when and how much data to send vs what to send.

Conservation of packets: New segment not injected into the network until and old segment has left

Self-clocking: The way TCP implements conservation of packets, i.e. segment transmissions are triggered by returning acknowledgements

Congestion avoidance: The equilibrium state algorithm for TCP

Slow-start: The algorithm which TCP uses to reach the equilibrium state

SACK implementation:

  • Fast retransmit and fast recovery are not resending already SACK’ed segments

Forward-most data: The highest seqnum that is correctly received


  • snd.fack = Forward most ACK
  • retran_data = size of outstanding retransmitted data
  • New ACK: snd.fack = snd.una = ACK’ed sequence number
  • DupACK: snd.una = ACK’ed sequence number, but snd.fack = highest seq in SACK
  • available window: awnd = snd.nxt - snd.fack + retran_data
  • Send new data whenever awnd < cwnd
    • If send new data, advanced snd.nxt
    • If retransmitting old data, increases retran_data
  • An incoming ACK: either decreases retran_data or advance snd.fack
  • If ACK advances snd.fack beyond snd.nxt at the time a segment was retx, the retx is lost

FACK start recovery when reassembly queue is longer than 3 MSS, i.e.

if ((snd.ack - snd.una) > (3*MSS) || (dupacks == 3)) {

Recovery ends when snd.una advanced beyond snd.nxt at the time the first lost is detected During recovery, cwnd held constant

Bibliographic data

   title = "Forward Acknowledgement: Refining TCP Congestion Control",
   author = "Matthew Mathis and Jamshid Mahdavi",
   journal = "ACM SIGCOMM Computer Communication Review",
   volume = "26",
   number = "4",
   month = "Oct",
   year = "1996",