Congestion Control for High Bandwidth-Delay Product Networks

This paper suggests that the inevitable move towards networks with high bandwidth-delay products will cripple the congestion control mechanism currently provided by TCP. In its place, the authors propose a new "explicit" congestion control protocol, XCP. The key feature of XCP is that unlike TCP, it separates efficiency control and fairness control. The authors argue that since efficiency and fairness are independent factors, decoupling them provides optimal control. For example, AIMD achieves fairness but the additive increase isn't quite the most efficient solution for high-bandwidth connections as too many RTTs are wasted in achieving the capacity. XCP continues to use an AIMD scheme for fairness control but instead opts for MIMD for efficiency control - the separation allows each control to use a suitable scheme. Despite using the AIMD scheme employed by TCP, XCP has quicker convergence to fairness than TCP since multiplicative decrease is performed more often than merely on the detection of dropped packets. Upon describing the XCP control mechanisms, the authors present simulations that compare XCP and TCP in the presence of previously discussed features such as RED and CSFQ queues. The results demonstrate that XCP matches TCP under normal conditions and that it outperforms TCP when the per-flow delay-bandwidth product becomes large.

This paper was a surprise after reading so many older publications that assumed TCP congestion control. The arguments made for XCP seem quite convincing; that said, I'd be surprised if there are no tradeoffs in TCP's favor as this paper appears to imply. The provision for security extensions is compelling and the section on XCP's TCP-compatibility is a very good idea considering the resistance to the status quo that always arises to some degree in technology. Papers like this really diversify the syllabus and differentiate this course from the standard undergraduate networking course.