Why Don’t Delay-based Congestion Estimators Work in the Real-world?

A number of designs have been proposed for complementing TCP’s treatment of packet loss as an implicit signal of congestion, with a signal derived from measurements of round-trip times (RTT). The premise of such delay-based congestion estimators (DBCEs) is that congestion is reflected in queueing delays that can be detected by measuring changes in RTT. We conduct a large-scale empirical analysis of real-world TCP connections to evaluate the effectiveness and limitations of five prominent DBCEs. Our findings are that none of the five perform well (correctly indicate congestion before a loss is experienced) for a large percentage of real-world TCP connections. They also often perform poorly by having high rates of false-positive and false-negative estimates of congestion. Further, we find that the connection characteristics that most influence the performance of these DBCEs are so diverse that designing an effective DBCE for all types of connections is still an open research problem.1