TCP with faster recovery

Among the problems affecting the current version of TCP are the slow recovery upon a coarse timeout expiration on long, fat pipes, and the reaction to random segment losses. Both problems are known to reduce the throughput of a connection. We propose and evaluate the merits of a class of TCP modifications obtained through a source-based estimate of the available bandwidth by measuring the rate of received ACKs. The estimated bandwidth is used to set the slow start threshold and the congestion window after a timeout or 3 duplicate ACKs. The goal is to allow sources to recover quickly after sporadic losses over high bandwidth-delay links. It is worth noting that only a slight modification of the protocol stack at the source is needed. In our algorithms, a TCP source estimates the bandwidth available to it using an exponential averaging. Whenever an ACK is received, the bandwidth estimate is updated based on the amount of data that clears the transmission buffer following the ACK reception, divided by the current RTT estimate. After a timeout or 3 duplicate ACKs, the available bandwidth estimate is used to reset the TCP congestion window and the slow start threshold. Simulation results show that TCP with "faster recovery" exhibits higher goodput than other flavors of TCP, notably TCP Reno and TCP SACK (selective acknowledgement) in specific scenarios.