Congestion Control for High Bandwidth-Delay Product Networks

D. Katabi, M. Handley, C. Rohrs, "Congestion Control for High Bandwidth-Delay Product Networks," ACM SIGCOMM Conference, (August 2002)

This paper presented eXplicit Control Protocol (XCP) as a replacement to the current widely deployed Transport Control Protocol (TCP) which introduced the new concept of decoupling utilization control and fairness control. The advantages of XCP (and disadvantages of TCP) were:

1. Unlike TCP where efficiency and fairness are coupled together into AIMD, XCP allows these two things to be decoupled. As a result, XCP effectively implemented MIMD for its efficiency controller and AIMD for its fairness controller.

2. In XCP, each packet sends a congestion header to the gateway which contains the sender's current window size and its RTT. The router in return gives back its feedback in the same header. This allowed that the new protocol does not maintain any per flow state in routers.

3. TCP's additive increase policy makes it less responsive in its ability to acquire spare bandwidth.

4. XCP facilitates the detection of misbehaving sources.

5. XCP provides a good incentive for users to deploy it. However, it can very well co-exist with existing TCP which allows it to be implemented gradually.

6. The parameters  used in XCP are pre-specified which makes it easy to be deployed without much application-specific tuning.


The design of XCP was followed by extensive simulations and stability analysis. The authors showed that XCP clearly outperformed TCP over RED, CSFQ, REM and AVQ.  Overall, this paper being a relatively new paper deals with some of the core issues that affect the Internet today and questioned the very assumptions of TCP. I really liked this approach of stepping back a bit and questioning the core assumption of why should packet loss be a indicator for congestion! Moreover, many of the ideas mentioned in this paper built on things that we had been discussing in the last few classes involving burst traffic and high delay and bandwidth links. Moreover,  I felt that for identifying misbehaving hosts, we do have keep per-flow information, which was not really explained well in the paper. To conclude, it would be really interesting to discuss the applicability of this protocol and how successful was its gradual deployment.