Measuring the Congestion Responsiveness of Internet Traffic

A TCP flow is congestion responsive because it reduces its send window upon the appearance of congestion. An aggregate of nonpersistent TCP flows, however, may not be congestion responsive, depending on whether the flow (or session) arrival process reacts to congestion or not. In this paper, we describe a methodology for the passive estimation of traffic congestion responsiveness. The methodology aims to classify every TCP session as either "open-loop" or "closed-loop". In the closed-loop model, the arrival of a session depends on the completion of the previous session from the same user. When the network is congested, the arrival of a new session from that user is delayed. On the other hand, in the openloop model, TCP sessions arrive independently of previous sessions from the same user. The aggregate traffic that the open-loop model generates is not congestion responsive, despite the fact that each individual flow in the aggregate is congestion responsive. Our measurements at a dozen of access and core links show that more than 60-80% of the traffic that we could analyze (mostly HTTP traffic) follows the closed-loop model.