Congestion Control for High Bandwidth-Delay Product Networks

EXplicit Control Protocol, XCP, is introduced in this paper as a bottom-up redesign of congestion control without the limitations of TCP. By allowing a better feedback mechanism than dropping packets (in fact making the feedback an explicit part of the underlying protocol, and not dropping packets at all), XCP better communicates both over- and under-utilization of resources to clients by modifying packet headers at intermediate gateways and having clients react based on the feedback. Separating out efficiency and fairness issues into two different controllers, an XCP gateway attempts to gain the best efficiency in a fair manner; the EC decisions are implemented by feedback given by the FC.

The best summarized explanation of the two controllers is that the EC uses Multiple-Increase Multiple-Decrease to rapidly hit efficiency targets, and the usual AIMD to converge to fairness. Most interestingly, the parameters for XCP can be set once for all workloads, and do not depend on the kinds of traffic the gateway receives, unlike RED.

The performance of XCP is explored in several simulations, and it achieves extremely high efficiency for any number of flows, as well as good fairness--- always much fairer than TCP. The one downside in the performance section is that it seems to be a bit less efficient than some of the other algorithms; the paper says it "trades a few percent of utilization for a considerably smaller queue size" but I fail to see the advantage of having smaller queue sizes if utilization is not being fully reached.

Of course, XCP is not TCP and integrating the two is covered in the paper to some extent. Gradual deployment would require gateways that understand both protocols and performed some fairness calculations between the two; this intermediate level of implementation seems much too complex, but unavoidable given TCP's hold.

The paper shows how congestion control "should" have been designed into routers and the protocol, as opposed to how it was integrated after the fact into TCP due to the discovery of failings of the original underlying algorithm.