TCP uses a round-trip delay estimate for its adaptive windowing
scheme to transmit data reliably over an unreliable network with
time varying bandwidth. An estimate of the round trip time is
maintained as a smoothed moving average [6] using the relation -
Similarly a smoothed variance (estimated as mean difference to
avoid square root calculations in the kernel) is also maintained (Tahoe
TCP). A initial re-transmission timeout is then set as -
If an acknowledgement for a segment is not received within the timeout, it is re-transmitted.
TCP uses a congestion window in the sender side to do congestion avoidance. The congestion window indicates the maximum amount of data that can be sent out on a connection without being acknowledged. TCP detects congestion when it fails to receive an acknowledgement for a packet within the estimated timeout. In such a situation, it decreases the congestion window to one maximum segment size (MSS), and under other cases it increases the congestion window by one MSS. There also exists a congestion window threshold, which is set to half the congestion window size at the time when a re-transmit was required.
The inherent assumption in this mechanism is that lack of an acknowledgement is due to network congestion. If a packet, however, is lost by the network for reasons other than network congestion, then waiting for the timer to run out is wasteful. This is a situation that may happen quite frequently in wireless networks, and so to improve TCP performance, it is needed to pre-empt re-transmissions before waiting for the timer to run out.
To guard against this scenario, Reno TCP [7] uses Fast Re-transmit and Fast Recovery algorithms. Both these algorithms depend on counting duplicate acknowledgements sent by the data receiver in response to each additional segment received following some missing data. Fast Re-transmit detects loss of a segment when three duplicate acknowledgements are received, and re-transnits it. Fast Recovery algorithm attempts to estimate how much data is outstanding in the network by counting duplicate acknowledgements.