Research On Internet Congestion Control Algorithms

Due to the unbalance distributions of network resources and traffics, the occurrence of congestion is an intrinsic characteristic of the Internet. Congestion control is necessary to keep the stability of the Internet. Congestion control algorithms can be broadly classified into two categories: source algorithms and link algorithms. Link algorithm, such as Active Queue Management (AQM), is executed in network devices to detect congestion and generate feedback information. Source algorithm, such as TCP, is executed in end hosts or edge devices to adjust sending rate in response to feedback. According to the end-to-end argument, TCP is traditionally employed to execute most congestion control. With the development of Internet, many limitations of the traditional congestion contril mechnisams emerged. For example, the window decrease policy of TCP is not suitable for streaming media and TCP has significant performance degradation in high speed networks. To address these problems, several issues about congestion control is studied in this thesis, including source algorithms, decoupling efficiency and fairness control, explicit algorithms over hybrid links which consists of high speed and wireless links, modeling and stability analysis. The main research works are as follows.A gentle slow start algorithm using bandwidth measurements, called gentle slow start (GSS), is proposed. The slow start of TCP has two shortcomings: firstly, when a connection is established TCP sets a random value for slow start threshold (ssthresh); secondly, TCP's exponential window increase policy is too aggressive. GSS addresses these two problems. For a new connection, GSS estimates its attainable bandwidth and uses the bandwidth to set the initial value of ssthresh. During the slow start procedure caused by timeout, GSS holds an adaptive parameter. When the size of congestion window exceeds the value of the parameter, GSS gently increase the window size and can smoothly convert to congestion avoidance procedure.Under the assumption of synchronous feedback, several fairness and TCP-friendliness theories of unicast congestion control has been proved, based on the works of R. Sastry and S. Lam et al. These theories are in terms of window increase/decrease functions. By choosing different window change function, congestion control algorithms can be customized to suit different application. We demonstrate this by a stream media algorithm. Motivated by the idea of"cross-layer design", we present the concept of"customizing congestion control strategy"where the application layer can choose its congestion control algorithm and also discuss some implementation issues.The decoupling of fairness and efficiency control of XCP and VCP are fully analyzed. XCP executes fairness control (FC) and efficiency control (EC) simultaneously in the network devices and conveys much information between end systems and network. In each control interval, both EC and FC are executed. VCP executes fairness and efficiency control in the end systems and conveys little information between end systems and network. In each control interval, only one controller is executed. Furthermore, we point out that the few executing times of FC results in significant fairness gap between XCP and TCP+AQM style protocol, such as VCP. By increasing the executing times of FC, we propose Fair VCP (FVCP) algorithm which outperforms VCP in fairness.Several adaptive VCP (AVCP) algorithms are presented. By extending VCP, a protocol aimed at high speed links, to wireless links, AVCP can be adapted to high speed links, wireless link and hybrid of the two links. AVCP is an explicit algorithm. Its key idea is that the sender conjectures the reason of each packet loss, congestion induced or corruption induced, with the help of"path load factor"provided by network and then decides to reduce the congestion window size or not. For explicit algorithms, network needs link capacity to calculate the feedback information. However, it is difficult to estimate accurate capacity for wireless links and contention-based Ethernet etc. With a control theoretic analysis, we show that the capacity estimation error may disturb the explicit congestion control algorithms to achieve promised performance.From an optimization point of view, congestion control modeling and its stability is well studied. We first correct several errors of the Kelly model from the viewpoint of physical meanings. Based on the work of Zhang et al., we then propose a general model called General Kelly Johari Zhang model (GKJZ) that is independent of any idiographic algorithms. The GKJZ model uses N dimensions difference equations with heterogeneous delays to model the dynamics of congestion control. Under the framework of GKJZ, we model VCP and derive sufficient condition for its local stability based on this model. Different from most of the existing stability conditions, our stability criterion is independent of delay and has a very simple form which is easy to verify. With our stability criterion, the congestion control algorithm achieves robustness because its control parameters can be statically configured and do not need to change according the delay.