Research on Congestion Control

Receiver-oriented and Measurement-based Transmission Control
in Heterogeneous (Wired/Wireless) Networks

Chi Zhang, School of Computer Science, Florida International University
Vassilis Tsaoussidis, Dept. of Electrical & Computer Engineering, Demokritos University, Greece

Traditional transmission control of reliable protocols, as exemplified by TCP, is based on the Additive Increase / Multiplicative Decrease (AIMD) algorithm [1, 2]. AIMD is somewhat a "blind" mechanism in a sense that the congestion window increases steadily until congestion occurs, and congestion is detected by missing segments. In [5], we identified five distinct limitations of traditional congestion control for potential improvements, which have been discussed in the recent literature:

An inherent characteristic of a "blind" window increase strategy is the natural cause of congestion and the subsequent need for downward window adjustments.
Additive increase is not efficient when the network dynamics encompass rapid changes of bandwidth availability.
Multiplicative decrease by a factor of 2 leads to severe transmission gaps, which degrade the performance of real-time applications that experience delay jitter and throughput fluctuation .
In the presence of errors caused by random wireless interference, congestion-oriented downward adjustments are triggered unduly and the available bandwidth is under-utilized.
Congested or lossy reverse paths that carry the receiver-generated acknowledgments might downgrade the transmission rates and the bandwidth utilization on the forward path.

Therefore, our research on congestion control for the next-generation transport protocols advances mainly on two fronts: the application requirements (e.g. quality of service) and the characteristics of the underlying networks (e.g. wireless links). Our recent research efforts are summarized as follows:

1. The Interrelation of Smoothness and Responsiveness in Heterogeneous and Dynamic Environments

A new family of protocols, namely the TCP-friendly protocols, are designed with two competing objectives: (i) to achieve smooth transmission adjustments for real-time applications, which is accomplished by reducing the window decreasing factor during congestion, and (ii) to compete fairly with TCP flows, which is accomplished by moderating the window increasing rate based a TCP throughput equation [3]. We were particularly interested in a parameterized TCP-friendly protocol, GAIMD [6], which generalizes TCP by parameterizing the congestion window increase value α and decrease ratio β.

In [8, 10, 12],we investigated the interrelation of TCP smoothness and responsiveness by studying the combined dynamics of friendliness-oriented α/β tradeoff and the network conditions for achieving efficiency and smoothness. We categorized three classes of TCP-friendly TCP(α, β) protocols: (i) Standard TCP, that is TCP(1, ½); (ii) Responsive TCP with relatively low β and high α (iii) Smooth TCP with relatively high β and low α. We confirmed experimentally that, in general, smoothness and responsiveness constitute a tradeoff; however, we uncover undesirable dynamics of the protocols when the network is heterogeneous (wired/wireless) or the flow characteristics do not follow a prescribed and static behavior:

Smooth downward adjustments confine the protocol's capability to exploit resources that become available rapidly after a handoff in mobile networks, and embarrass the fair and efficient growth of incoming flows.
In the context of wireless networks with high error rate, the conservative behavior of smooth TCP degrades the protocol performance for both best-effort and real-time applications.
In a system where smooth and responsive protocols co-exist and the contention is high, the responsive protocols may be favored.
Throughput of smooth TCP may be greater than that of standard or responsive TCP in a network with high delay-bandwidth product.

Furthermore, we realized that the ultimate reasons behind the above phenomena and observations are some inherent characteristics of equation-based TCP-friendly protocols:

It takes several RTTs for a low α to pay back the bandwidth credit of a high β, since TCP-friendly protocols are based on an equation that gives the steady-state TCP throughput. In a dynamic environment where traffic load changes rapidly, friendliness might not be accomplished.
The TCP throughput equation [3] does not capture the system behavior when the network operates between the knee and the cliff defined in [2].
The current TCP throughput equation is inaccurate when the error rate is high and the responsiveness (i.e. α) dominates the protocol behavior.

We concluded that equation-based adjustments could guarantee neither efficiency nor friendliness on its own, in the context of heterogeneous networks and dynamic environments. A supportive mechanism for error classification and bandwidth detection may complement the equation-based adjustments and produce positive dynamics.

2. TCP-Real: Receiver-oriented and Measurement-based Congestion Control

We have developed TCP-Real [5, 7, 9], which employs a receiver-oriented and measurement-based congestion control mechanism that significantly improves TCP performance for both delay-sensitive and delay-tolerant applications in heterogeneous (wired/wireless) networks and over asymmetric paths (see Figures 2.1 and 2.2). TCP-Real relies on three novel congestion detection mechanisms beyond packet loss detection:

Receiver-oriented Congestion Detection : The receiver measures the network condition and attaches the results to the ACKs sent back to the sender.
Wave Pattern : In order for the receiver to observe accurately the level of contention and/or packet loss, a limpid communication pattern with the sender, namely Wave, is introduced in [4]. A wave consists of a number of fixed-sized data segments sent back-to-back, matching the inherent characteristic of TCP to send packets back-to-back.
Measurement-based Congestion Detection: The receiver computes the data-receiving rate of a wave, which is determined by the interleaving patterns of packets from different flows at the bottleneck router. The lower the perceived rate, the higher the level of contention suggested at the bottleneck, and vice versa. During congestion, the data-receiving rate might fluctuate dramatically due to packet drops that occur at the bottleneck router. However, if a packet drop is due to a wireless error, the data-receiving rate shall not be affected by the gap of missing packets, since the wave size is attached to the data packet as a TCP header option.

The novel congestion detection mechanisms distinguish the nature of a packet loss (due to congestion or transient wireless errors), and abrogate the impact of false assessments at the sender due to lost or delayed acknowledgements on a lossy reverse path. In addition, the level of contention is perceived before packet drops occur. Consequently, based on the more accurate information on the network condition, a sophisticated error recovery mechanism is enabled at the sender side:

Decrease Decision based on Error Natures : The congestion window is reduced only when a drop is associated with congestion. Therefore, fairness and TCP-friendliness are not violated, and the protocol performance is significantly improved in the presence of random wireless errors.

3. TCP(α, β, γ, δ): Improving TCP Smoothness by Synchronized and Measurement-based Congestion Avoidance

The challenge of improving TCP smoothness for real-time applications does not lie in simply achieving smooth window adjustments, but rather in providing smoothness along with bandwidth efficiency, fairness, responsiveness, as well as controlled queue length.

We observed in [11] that although multiplicative decrease is necessary to accomplish fairness in congestion control, it does not inevitably sacrifice system throughput, as long as the system operates between the knee and the cliff, according to an equation. However, even when the system throughput is relatively stable, end users of real-time applications do not necessarily experience smooth performances (see Figure 3.2). We emphasized that in a real system multiplicative window decreases are often unsynchronized among competing TCP flows, due to random congestive drops. We further argued that the major obstacle for achieving smoothness is the unsynchronized and random window control, and analyzed its negative impact on smoothness, short-term fairness, and even long-term fairness measured by a worst-case fairness index.

We therefore proposed an experimental congestion avoidance mechanism [11] to improve TCP smoothness for real-time applications. The mechanism relies on a fine-grained RTT estimation to measure the network condition, and coordinates the upward and downward window adjustments to abolish the damage of unsynchronized window control. Congestive packet drops are reduced by a new control parameter γ, which determines the window decreasing ratio when the level of contention exceeds a threshold that indicates an upcoming congestion. The bottleneck queue length can also be controlled in an end-to-end fashion. Simulation results confirm that the new mechanism enhances significantly the smoothness and fairness (see Figures 3.1, 3.2 and 3.3), without a cost of responsiveness. In fact, by enabling a new parameter δ, the responsiveness can be even enhanced adaptively when the bandwidth is under-utilized.

References

1. M. Allman, V. Paxson and W. Stevens, "TCP Congestion Control", RFC2581, April 1999.
2. D.-M. Chiu and R. Jain, "Analysis of the Increase and Decrease Algorithms for Congestion Avoidance in Computer Networks", Computer Networks and ISDN Systems, 17(1):1-14, June 1989.
3. J. Padhye, V. Firoiu, D. Towsley, and J. Kurose, "Modeling TCP Throughput: A Simple Model and its Empirical Validation", In Proceedings of ACM SIGCOMM '98, August 1998.
4. V. Tsaoussidis, A. Lahanas and C. Zhang , "The Wave & Probe Communication Mechanisms", Journal of Supercomputing (Kluwer), 20(2):115-135, September 2001. PDF
5. V. Tsaoussidis and C. Zhang , "TCP-Real: Receiver-Oriented Congestion Control", Journal of Computer Networks (Elsevier), Vol. 40, No. 4, November 2002. PDF
6. Y.R. Yang and S.S. Lam, "General AIMD Congestion Control", In Proceedings of the 8th International Conference on Network Protocols, November 2000.
7. C. Zhang and V. Tsaoussidis , "Exploiting the Potential of TCP-Real in Multiplexed Wired/Wireless Channels", The 1^st International Workshop on Wired/Wireless Internet Communications (WWIC 2002), June 2002.
8. C. Zhang and V. Tasoussidis , "On the Efficiency of TCP-friendly Window Adjustment Strategies in Wired/Wireless Networks" (invited paper), Conference on Scalability and Traffic Control in IP Networks, SPIE ITCOM 2002, July 2002.
9. C. Zhang and V. Tsaoussidis , "TCP-Real Improving Real-time Capabilities of TCP over Heterogeneous Networks", The 11^th International Workshop on Network and Operating Systems Support for Digital Audio and Video (ACM NOSSDAV 2001), June 2001. PDF
10. C. Zhang and V. Tsaoussidis, "The Interrelation of TCP Smoothness and Responsiveness in Heterogeneous Networks", The 7^th IEEE Symposium on Computers and Communications (ISCC 2002), July 2002. (Runner-up award winner) PDF
11. C. Zhang and V. Tsaoussdis, "Improving TCP Smoothness by Synchronized and Measurement-based Congestion Avoidance", The 13^th International Workshop on Network and Operating Systems Support for Digital Audio and Video (ACM NOSSDAV 2003), June 2003. PDF
12. V. Tsaoussidis and C. Zhang, "The Dynamics of Responsiveness and Smoothness in Heterogeneous Networks", IEEE Journal on Selected Areas in Communications, March 2005.