COOPERATIVE CONGESTION CONTROL SCHEMES IN ATM NETWORKS

Traffic and congestion control is a major issue in ATM networks. One o f the basic problems faced in the design of efficient traffic and cong estion control schemes is related to the wide variety of services with different traffic characteristics and quality of service (QoS) requir ements supported by ATM networks. In this article, the authors propose a new way of organizing the control system so that complexity is easi er to manage. The multi-agent system approach, which provides the use of adaptative and intelligent agents, is investigated. This is a power ful way to introduce a degree of intelligence into an otherwise purely algorithmic approach. The authors show, through the two congestion co ntrol schemes proposed, how to take advantage of using intelligent age nts to increase the efficiency of the control scheme. First, TRAC (thr eshold based algorithm for control) is proposed, which is based on the use of fixed thresholds which enables the anticipation of congestion. This mechanism is compared with the well-known push-out algorithm and it is shown that the authors' proposal improves performance. Also dis cussed is the necessity of taking into account the dynamicness of the network and investigating the multi-agent system approach. In TRAC, ad aptative agents with learning capabilities are used to tune the values of the thresholds according to the status of the system. However, in this scheme, when congestion occurs, the actions we perform are indepe ndent of the nature of the traffic. Subsequently, we propose PATRAC (p redictive agents in a threshold based algorithm for control) in which different actions are achieved according to the QoS requirements and t o the prediction of traffic made by the agents. Specifically, re-routi ng is performed when congestion is heavy or is expected to be heavy an d the traffic is cell loss sensitive. This re-routing has to deflect t he traffic away from the congestion point. in this scheme, we propose a cooperative and predictive control scheme provided by a multi-agent system that is built in to each node.