Dynamic Behavior of Slowly-Responsive Congestion Control Algorithms 

Deepak Bansal, Hari BalaKrishna, Sally Floyd and Scott Shenker 

Presented By: Poovappa Subbaiah (Dinu)

Summary: 

Applications such as best-effort, unicast streaming video and audio require steady transmission rates for optimal performance with the available bandwidth. Based on the TCP-Compatible paradigm, which requires that any mechanism have the same bandwidth usage as TCP in the presence of constant loss rate, several slowly responsive congestion control algorithm have been proposed. By responding less drastically than the AIMD, these algorithms attempt to provide a smooth data rate curve. While the proposed algorithms are TCP-Compatible in static state, this paper attempts to determine its behavior in a dynamic situation. 

Several TCP-Compatible algorithms are considered such as Binomial, TEAR, TFRC and AIMD (b < 0.5) for testing. The dynamic test scenarios present some interesting behavior of these algorithms. A plot of the stabilization time as a function of the algorithm parameter shows that TFRC (self clocked) and SQRT respond extremely well in comparison to TCP. A test for long-term fairness reveals that the TFRC and SQRT do not compete unfairly with TCP in the presence of varying bandwidths. It is further determined that there is a loss in throughput for the slowly responsive mechanisms in comparison to TCP in the same environments. A test for smoothness of data rates in dynamic conditions show that TFRC achieves the required property. Based on all these factors, the paper concludes that most of the TCP-Compatible algorithms studied are safe for deployment. An important contribution of this paper is to show that incorporating a self-clocking mechanism based on packet conservation prevents the network from persistent loss rates even with the more slowly responsive algorithms. 

Group Discussions: 

The analysis of the paper resulted in several interesting questions and discussions. 

• One of the questions that were discussed was if the tests conducted were too dynamic. Since we use optical cable for data transmissions, data rate fluctuations on wired lines are fairly smooth and packet drops are minimal. Would this kind of extreme dynamic testing ignore the real internet and fail in identifying the correct behavior of the slowly responsive mechanisms.  Contrary to this, was the argument that the tests conducted were probably not dynamic enough. In support, it was suggested that a single bottleneck, used in the tests in the paper did not emulate the real environment.
• Would it be possible to emulate the slow start with the TFRC? While the sudden response to congestion needs to be slow, we also need a faster method to utilize available bandwidth. This would prevent the slowly responsive mechanisms to lose bandwidth to TCP. This question is being discussed within the Network community and in fact, when the data rate reaches a certain low point, slow start is used to push up TFRC data rates. However, it is surprising that there is no mention of it in the paper.
• Instead of suggesting a new slowly responsive algorithm, wouldn’t it be better to slow down TCP congestion control algorithm itself? This would minimize the concerns about deploy ability. It is possible to slow down TCP congestion control algorithm and in fact it is also discussed in the paper. The AIMD using a decrease parameter, b as lesser than 0.5 is slowly responsive, The paper also discusses the Binomial algorithms and several tests are conducted on the SQRT algorithm which is a non linear generalization of the AIMD.
• Does queuing algorithms like RED help in reducing the packet losses in the slowly responsive algorithms? RED tries avoid congestion by controlling the average queue size at the gateway or router. First problem when network gets congested is that router's queue overflows and begins to drop packets. This situation can be prevented in the sender reduces the sending rates before the congestion happens and slower responsive mechanisms could be benefited by this. But unfortunately the performance of RED itself is in question. By and large the network community considers RED to have been unsuccessful in its performance.
• What about the sensitivity of smoothness? How sensitive must it be or in other words how smooth must the data rate curve be to be reasonably good? The paper Equation-based Congestion Control for Unicast Applications by Sally Floyd, Mark Handley, Jitendra Padhye and Jörg Widmer discusses this aspect briefly. The authors suggest that 0.15 secs interval seems to be a plausible candidate for a minimum interval over which bandwidth variations would begin to be noticeable to multimedia users. However this is very relative and depends on different aspects, including the content of the stream itself.
• A couple of observations that were made in the discussions require mention. The basic assumption that multimedia streams need a smooth data transfer rates could be questionable. This comes in relation to the fact that many variable bit-encoding schemes have been proposed and have been found to perform better than a constant rate encoding. Also, The smoothness of the curve is a property of the intervals in time that the rates are averaged for. This could be a highly fluctuation curve with a high frequency of oscillations, which can result in a fairly smooth curve, for a certain time interval of measurement. This is an interesting point and possibly needs further study.

Additional References: 

• The following paper discusses the Binomial congestion control algorithms in greater detail. As seen in the tests for TCP Compatibility in a dynamic environment, the SQRT performs extremely well.

              Bansal, D., and Balakrishnan, H. Binomial Congestion Control Algorithms. In Proceedings of the Conference
              on Computer Communications (IEEE Infocom) (A nchorage, AK, April 2001). Pp. 631-640.

• The official website for TEAR: www.csc.ncsu.edu.

Click here for Presentation Slides