The authors give a new algorithm for calculating the retransmit times (smoothed round-trip time, SRTT, and retransmission time-out, RTO) in a TCP network. The problem with current methods is that acknowledgments do not indicate whether they are for retransmissions or for the packet originally sent. This ambiguity can throw off estimates of SRTT and RTO, causing the retransmit timers to be wildly off.
The authors review several current methods for correcting this problem, giving reasons why some of these methods cannot correct the problem and may make things worse. The problem could be solved by serializing acknowledgments so that they can be matched to the transmissions to which they are responding.
The authors suggest that retransmission times (SRTT and RTO) only be calculated based on acknowledgments for packets that have not been retransmitted. They further suggest that back-off on RTO not occur until after a retransmission attempt has been made and has failed. Using a statistical argument, the authors give an upper bound on RTO based on prior estimates of SRTT. One must be careful in choosing the method used to calculate RTO and SRTT to ensure a quick and stable convergence, something that is not guaranteed by any of the proposed algorithms.
The authors give some good suggestions and make some valid observations. It would have been better for their case if they had presented some data. A worthwhile project would be gathering data on retransmission times on IP networks in order to determine the effect of different algorithms for calculating SRTT and RTO on throughput and network efficiency. The authors leave open the problem of balancing the problems of throughput and transmission delays that are common in IP networks.