Congestion Control in Data Transmission Networks

In recent years, we have experienced rapid evolution of networking services. We can find widespread network traffic related to web browsing, e-mail, electronic trade and money transfer, telematics (vehicle positioning, crash notification, etc.), video streaming, remote visualization and steering, Internet telephony, etc. As a consequence, in addition to the increased intensity, the today networks need to handle also the diversity of data streams and meet their QoS requirements. Together with serving the applications demanding fast and reliable information interchange (e.g., banking transactions), modern data transmission networks are expected to provide high-throughput, low-jitter, end-to-end connectivity important for multimedia transmission. The connectivity needs to be available to both local and long-distance streams without violating fairness constraints. In the following sections, we will recall the fundamental concepts of organizing data transfer in communication networks and elaborate on a major obstacle to obtaining appropriate QoS – the congestion.

The congestion occurs when the traffic generated by the network users exceeds the available bandwidth in the communication system. In such circumstances, not all the packets sent by the sources can be immediately relayed on the route towards their destination. Instead, they accumulate in the buffers at the intermediate nodes and wait for the bandwidth increase. If the incoming rate is not reduced (or stopped) before the queue of awaiting packets reaches its limit, typically defined by the amount of the reserved memory at the node, the new data pieces must be discarded. The lost fragments are retransmitted, which further deepens the congestion at the bottleneck point. At certain stage, the network becomes clogged with retransmissions and stops providing its services – this state is referred to as a deadlock or congestion collapse. In fact, the early communication networks frequently suffered from congestion collapse, until the development of the Jacobson’s scheme [73] for the Internet flow control.

The main purpose of control engineering is to steer the regulated plant in such a way that it operates in a required manner. The desirable performance of the plant should be obtained despite the unpredictable influence of the environment on all parts of the control system, including the plant itself, and no matter if the system designer knows precisely all the parameters of the plant. Even though parameters may change with time, load, and external circumstances, still, the system should preserve its nominal properties and ensure the required behavior of the plant. In other words, the principal objective of control engineering is to design control (or regulation) systems which are robust with respect to external disturbances and modeling uncertainty. This objective may be very well achieved using the sliding-mode technique [6, 11, 16, 18, 26, 28, 31, 43, 45, 55, 62, 66, 78, 79, 81, 85, 87], which is extensively used throughout this monograph. To be more precise, in the monograph, we focus our attention on the application of discrete sliding-mode control principles to the congestion elimination in data transmission networks. However, in order to make the text self-contained, we begin this chapter with presenting the main notions and concepts used in the field of variable structure systems and sliding-mode control.

In this chapter, we introduce the fundamental concepts behind the fluid-flow modeling of data traffic in communication networks. We emphasize the effects caused by action delay, which is the time that elapses from the moment the control information (or the controller command) is sent by a network node, the information reaches the data source it is destined for, appropriate action is taken by the source, and until subsequently that action affects the state of the node that issued the command. Indeed, as recognized in many significant papers, for example [3, 6, 7, 10, 14, 17, 22, 26], the existence of action delay constitutes the main obstacle in providing efficient control in data transmission networks, and it should be explicitly considered in the controller design and system analysis. Since we intend to make use of the benefits of SMC, which is well known to be robust and efficient regulation technique successfully applied in many engineering areas (see, e.g., recent special issues [4, 15, 25]), it is of paramount importance to account for the adverse effects of delay. This is due to the fact that delay reduces the system robustness – typically, mismatched perturbations are introduced and the invariance property [9] no longer holds – which threatens stability of the sliding motion. In the design procedures presented in this chapter, we overcome the delay problem by an appropriate selection of the switching function which incorporates a state predictor. In this way, the delay in the feedback loop no longer poses a stability threat, and the system dynamics can be tuned for the maximum responsiveness to the changes of networking conditions reflected in the fluctuations of the available bandwidth.

In this chapter, we direct our attention to the design of flow control algorithms for networks in which the feedback information about the current network state is accessible for source rate adaptation at discrete time instants only. In this type of networks, in addition to the effects of nonnegligible delay, the design procedures need to explicitly account for the phenomena related to finite sampling rate. Hence, in this chapter, both the modeling and the controller design are performed directly in discrete-time domain.

In Chap. 5, we analyzed the basic networking phenomena related to controlling the flow of data in the network in which the feedback information about the network state is accessible for input rate adaptation at discrete-time instants only. We considered the problem of regulating the transmission rate in a single connection. In this chapter, we study a more complex setting in which the controller placed at the bottleneck node regulates the flow of data in multiple connections originating at various sources. It is assumed that the controlled connections can be characterized by different round-trip times, which is a typical situation in real networks. Consequently, the model developed in Sect. 5.1 for the case of a single-flow needs to be extended to cover the case of simultaneous transmission rate control in multiple connections.

For telecommunication operators, the cost of running a particular algorithm is an important factor when deciding about its applicability in the supervised network. Therefore, the business point of view will generally favor such control strategies which allow for the explicit specification (or at least estimation) of the amount of the exchanged feedback information in relation to the transferred users’ data. In consequence, emitting feedback information carriers at regular time intervals (which is assumed in traditional approaches for discrete-time network modeling, e.g., [2–5]) is not cost efficient in the economical terms. Instead, one can send a feedback information carrier every N data packets and, in this way, place a direct limit on the extent of the transmitted management traffic with respect to the profit generating transmission of the users’ data. Since this method relies neither on continuous feedback information availability nor on maintaining the synchronization of constant sampling period (which is a serious challenge in multisource systems [1]), it is more scalable and requires less control effort than the classical regulation schemes presented in the literature. However, sending a control unit every N data packets, and not every T seconds, means that (after RTT) the feedback information will be available for rate adaptation at the sources at irregular time instants. Consequently, in order to maintain adequate system performance, the variable, input-dependent sampling period should be explicitly accounted for in the design of flow control algorithm. We will show, however, that the controllers developed for the system with constant discretization period in Chaps.​ 5 and 6 can be quite intuitively adapted for the case of variable sampling rate analyzed here. Moreover, we will demonstrate that provided that certain additional constraints are met, the new strategies maintain the favorable properties of constant-sampling-rate controllers. In particular, the proposed schemes will be shown to eliminate packet losses originating from unknown bandwidth variations and to ensure full bandwidth usage.

Before introducing the proposal of Van Jacobson [9], the TCP/IP-based networks suffered from severe congestion problems. The bursts of traffic intensity frequently led to a network breakdown called the congestion collapse, and the resulting throughput degradation by several orders of magnitude. The Jacobson’s algorithm, implemented at the connection end points, ensured the basic control mechanism used to regulate the amount of data injected into the network. According to this algorithm, the transfer rate of a TCP source (or more specifically the window size) is increased until the congestion is detected at some link in the network. Initially, the window size at the source is enlarged by the number of packets acknowledged by the receiver. It is called the slow-start, or exponential-growth phase, and is used to quickly capture enough bandwidth to transmit the user’s data at a sufficiently fast rate. When a certain threshold value is reached, ssthresh, the window size continues to grow, but at a slower rate. In this phase, called the congestion avoidance, the window is enlarged by one packet every RTT. The transmitter tries to reduce the risk of link buffer overflow at the remote node(s), and the window size increases approximately linearly in time. Usually, the bulk of the user’s data is transmitted in this phase (see, e.g., the analysis performed in [19]).

Watching how ubiquitous communication networks have become in recent years, there is no question about their importance for social life and modern economy. They provide means of information sharing and data exchange basically anywhere, anytime. However, the numerous advantages of being able to quickly communicate thoughts and ideas over large distances, and gain access to remote resources, can only be achieved if the communication system is administered in a proper way. A major impact on the condition of a network is attributed to the mechanisms of control. They should allow efficient resource sharing according to the set of constraints specified for a given network (such as the channel capacity). More importantly, however, the implemented control methods should provide appropriate reaction to the dynamically changing networking conditions. In this way, one may avoid bottlenecks, loss of data, and guarantee adequate service level.