Using signaling methods to improve the performance of TCP congestion control.

摘要

In the TCP congestion control problem, TCP senders adjust their transmission rates in response to (a possible lack of) congestion signals received from the IP routers along their transmission path. There has been considerable research into the design and performance evaluation of both Active Queue Management (AQM) algorithms and TCP congestion avoidance algorithms. However, very little attention has been paid to the study of congestion signaling methods, which couple the AQM algorithms with the TCP congestion avoidance algorithms over the network. By focusing our attention on this topic, in this dissertation we show that congestion signaling methods can also have a significant effect on TCP congestion control.; First, we propose a new congestion signaling method called ACK Spoofing, and demonstrate through detailed simulation experiments that it outperforms all other signaling methods including packet dropping and Explicit Congestion Notification (ECN). The basic idea of ACK Spoofing is that an IP router can force a TCP sender to execute the TCP fast retransmit/fast recovery algorithms and reduce its transmission rate, by sending it multiple artificially-generated duplicate acknowledgments. This method shortens the congestion signaling path in comparison to ECN, which improves the effectiveness of existing AQM algorithms. We also study its implementation complexity and design a number of enhancement mechanisms, including the on-demand state maintenance scheme.; Second, we introduce a new concept called Congestion Signal Cancellation, which can be combined with either the ACK Spoofing or ECN signaling methods. With congestion signal cancellation, we give the router a second chance to examine the state of its queue at some time after the AQM algorithm originally generated the congestion signal. If conditions at the router have been improved sufficiently, then delivery of a now-obsolete congestion signal to the targeted TCP sender would be counterproductive, leading to degraded performance for the TCP target and/or underutilization of the bottleneck link. Thus, we define two methods by which a router can annul some of its previously-generated congestion signals from the network. Using modeling analysis and simulation, we demonstrate that adding congestion signal cancellation to existing AQM algorithms can yield remarkable performance improvements in terms of goodput and packet loss rate. In addition, congestion signal cancellation significantly reduces the RED algorithm's sensitivity to its parameter settings and enhances the system stability.