TCP-friendly exponential congestion control for multimedia communication

Congestion control is becoming an important issue in a shared network due to the growth of the network and the load. Congestion conk01 is concemed with allocating the network resowes and is heavily influenced by the traffic dynamic. It is not easy to get a satisfactory solution due to the complexity of the problem There is a number of network policies that affect the congestion control and there are several requirements for the design of a congestion control scheme. Among the existing congestion control algorithms, additive increase multiplicative decrease (m) a lgor ih [2,3] is widely deployed. This algorithm is the base of congestion control for Transmission Control Protocol (TCP)[22]. The key principle of AIMD is that it raises load by a constant if m loss occurs and responds to packet losses, i.e. congestion indication, by reducing load to a fraction of the current value. TCP is not well-suited for new applications such as streaming multimedia. To address this issue, the recent approach is to promote the use of end-to-end congestion contml schemes [6]. Since TCP is the dominant traffic, new congestion control schemes should be TCP-biendly. Thal means, the throughput of a non-TCP flow should be approximately the same as that of a TCP flow under the same conditions of round trip time and packet loss rate. However, it was recently shown that the additive increase nature of AIMD appears inferior [19]. Two desired properties of any congestion control scheme are fairness and convergence. But additive increase does not.guarantee the quickest convergence and AIMD is unfair under ceriain conditions, as pointed out in [19,3,10,11]. It has heen also shown in 1141 that a better smoothness can be achieved with a decrease parameter less than %. Moreover, large rate variations can be observed when loss rate is high. Other works [17,15,13,12] argued that increase-by-one decreaseto-half of AIMD 'is not a hndamental requirement of congestion control. Other values have been suggested for increase and decrease. One question arises, how to choose the increase and decrease parameters SO that high thughput can be achieved while obtaining the reasonable response time according to the dynamic behavior of the network. We propose therefore in this paper a general version for congestion control with a high order model. Instead ofusing concrete increase-by-one decrease-to-half such as in existing schemes, we use general parameters a for increase and b for decrease, respectively. Moreover, we consider the exponential behavior of the round-trip delay using an exponential increase function. With the model, we derive the throughput as a function of a and b, and of loss rate and roundhp delay. In addition, we obtain the relationship between a and b and show how to choose these parameters to achieve high throughput. The paper is organized as follows. In section 2 we review the basic principle of congestion control and some related works. We describe our genml version of congestion conkol in section 3. Section 4 shows OUT simulation results and finally, section 5 concludes the paper.