Improving Internet Congestion Control and Queue Management Algorithms


1. TCP over Diff-serv

This work demonstrates that with simple modifications to the network and end-hosts, bandwidth guarantees can be provided over the Internet using TCP and simple priority marking.  Specifically, the mechanism employed in the networks is an Enhanced RED queue. This queue management assumes that connections with reservations police their connections using a token bucket and that they mark packets [5] which are compliant.  The queue management mechanism gives preferential treatment to marked packets in that their drop probabilities are significantly smaller than those of unmarked packets.  Given this network support, we show that with minor changes in the end-hosts, we can effectively support bandwidth requirements in the network.  The work focuses on changes in the TCP sender, however, similar modifications can be easily applied to a UDP/RTP sender.  

The simulations were run using a modified version of ns-1.1.   Experiment files which were used to run the simulations are included underneath the graphs which display the results.  If you have any problems with the code, you can e-mail me.  The TCP mods and ERED are currently being implemented at IBM.

Finally, the slides I meant to show at NOSSDAV (before the laptop projector died on me) [1] are below.

[1] W. Feng, D. Kandlur, D. Saha, K. Shin, "Understanding TCP Dynamics in an Integrated Services Internet" NOSSDAV '97, May 1997. (Paper ps | Slides html | Slides ppt)

[2] W. Feng, D. Kandlur, D. Saha, K. Shin, "Understanding and Improving TCP Performance over Networks with Minimum Rate Guarantees" IEEE/ACM Transactions on Networking, Vol. 7, No. 2, pp. 173-187, April 1999. (Extended version of [1]) ps

[3] W. Feng, D. Kandlur, D. Saha, K. Shin, "TCP Enhancements for an Integrated Services Internet," IBM Research Report RC 20618, version 1, November 1996. ps  Note: This report contains additional experiments related to burst losses and the delayed send work. These simulations were removed from subsequent papers due to space limitations.

[4] D. Clark, "Adding Service Discrimination to the Internet" September 1995.


2. Adaptive packet marking for TCP

The work presents several adaptive packet marking mechanisms for providing soft bandwidth guarantees to individual flows or flow groups.   The schemes use modest support from the network in the form of priority handling for appropriately marked packets and rely on intellligent transmission control mechanisms at the edges of the network to achieve the desired throughput levels.   In our model, the user specifies a desired minimum service rate for a connection or a connection group. At any point of time, the sender, in cooperation with the packet marking engine, tries to achieve and potentially exceed the requested minimum service rate without using the high priority service. If however, it fails to achieve the minimum target rate, the packet marking engine starts prioritizing packets until the observed service rate reaches the desired target rate. Once the target is reached, it strives to reduce the number of priority packets without falling below the target. When the number of priority packets comes down to zero, the source tries to increase its share of the best-effort bandwidth.

We consider marking mechanisms of two distinct flavors: (1) where packet marking is completely independent of and transparent to the traffic source, and (2) where the marking module is integrated with the transmission control mechanisms at the source. To address the robustness of the algorithms, we evaluate their performance in situations where the network is over-subscribed and when the network contains non-responsive flows. We also consider scenarios where only some parts of the network support service differentiation and propose mechanisms for detecting and reacting to such situations at the source. Finally, we discuss issues in interoperating ToS-based mechanisms with alternative mechanisms for supporting service differentiation in the Internet. The simulations were run using a modified version of ns. All experiment files and source code will be made available soon.

[1] W. Feng, D. Kandlur, D. Saha, K. Shin, "Adaptive Packet Marking for Providing Differentiated Services in the Internet" U. Michigan CSE-TR-347-97 and IBM RC 21013,  October 1997.

[2] W. Feng, D. Kandlur, D. Saha, K. Shin, "Adaptive Packet Marking for Providing Differentiated Services in the Internet"  Proc. of 1998 International Conference on Network Protocols (ICNP '98),  October 1998. (Also IBM RC 21013 Version 2,  January 1998.) (Paper ps | Slides ps ppt)

[3] W. Feng, ns Modifications tgz.

[4] W. Feng, D. Kandlur, D. Saha, K. Shin, "Adaptive Packet Marking for Maintaining End-to-End Throughput in a Differentiated Services Internet", IEEE/ACM Transactions on Networking, v.7, n.5, Oct. 1999. ps


3. Adaptive RED and SubTCP

This work describes the use of Adaptive RED and SubTCP for effectively eliminating packet loss in congested TCP/IP networks.  The results show that one can reduce packet loss considerably by making active queue management techniques more cognizant of the offered load and by designing end-host congestion control mechanisms more intelligently.  When both techniques are combined, they form a synergistic combination which can achieve extremely high efficiency even in the most difficult scenarios.

[1] W. Feng, D. Kandlur, D. Saha, K. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks" U. Michigan CSE-TR-349-97,  November 1997. ps

[2] W. Feng, D. Kandlur, D. Saha, K. Shin, "A Self-Configuring RED Gateway", INFOCOM '99, March 1999. (This is a subset of the work in [1].) (Paper ps | Slides ppt)


4. Blue AQM

This work describes Blue, a fundamentally different queue management algorithm which can effectively eliminate packet loss in congested TCP/IP networks.  The results show that one can reduce packet loss considerably by decoupling congestion management algorithms from either the instantaneous or average queue length. ns-1.1 simulation modifications can be downloaded from [2]. The FreeBSD/ALTQ implementation can be downloaded from [3]. The algorithm has been implemented and validated by others [8,9,10] and was later acknowledged in an article on router bufferbloat by the pioneer of AQM [11].

[1] W. Feng, D. Kandlur, D. Saha, K. Shin, "Blue: A New Class of Active Queue Management Algorithms" U. Michigan CSE-TR-387-99,  April 1999. ps | pdf

[2] W. Feng "ns-1.1 modifications for implementing Blue and Stochastic Fair Blue" tgz

[3] W. Feng "ALTQ modifications for implementing Blue" (Program mods tgz | Kernel mods tgz)

[4] W. Feng, D. Kandlur, D. Saha, K. Shin, "Stochastic Fair Blue: A Queue Management Algorithm for Enforcing Fairness", in Proc. of INFOCOM 2001, April 2001. (Paper pdf | Slides ppt)

[5] W. Feng, D. Kandlur, D. Saha, K. Shin, "Blue: An Alternative Approach To Active Queue Management", in Proc. of NOSSDAV 2001, June 2001. (Paper ps | Slides ppt)

[6] W. Feng, D. Kandlur, D. Saha, K. Shin, "The Blue Queue Management Algorithms", IEEE/ACM Transactions on Networking, Vol. 10, No. 4, August 2002. pdf
Winner of the IEEE Communications Society 2003 William R. Bennett Prize for Best Paper in IEEE/ACM Transactions on Networking.
Also Winner of the Best IBM Research Paper Award in Computer Science, Electrical Engineering, and Math 2002.

[7] W. Feng, "Parameters for Blue", Unpublished. txt

[8] I. Bartok, "Implementation and Evaluation of The Blue Active Queue Management Algorithm", Diploma Thesis, Budapest University of Technology and Economics, May 2001. pdf

[9] S. Burri, "Blue: Active Queue Management", CS756 Project Report, George Mason University, May 2004. pdf

[10] E. Dumazet, "Patchwork [net-next-2.6,v3] net_sched: SFB flow scheduler", February 2011. Summary | Development notes

[11] K. Nichols, V. Jacobson, "A Modern AQM is Just One Piece of the Solution to Bufferbloat", ACM Queue, Vol. 10, No. 5, May 2012. paper


Follow-on work

1. RED and Multimedia Congestion Control

This work describes the bandwidth jitter that accompanies the use of randomized active queue management and its effect on the buffering requirements of streaming multimedia applications.

[1] W. Feng, W. Feng,  "The Impact of Active Queue Management on Multimedia Congestion Control," IC3N 1998, October 1998. (Paper ps | Slides ppt)