COST 257 TD(96) 10 September 1996 An enhanced batch model for resource allocation in ATM networks K. Sevilla, A. Gravey France Tt%com - CNET 1 Introduction One of the great ATM networks advantage is their flexibility to support different types of traffic. But on the other hand, resource allocation raises complex problems in order to ensure Quality of Service (QoS) requirements. Congestion control is essential and algorithmic procedures are developed for Connection Admission Control (CAC). If we consider an ATM node, traffic rules for allocating resources on this node to an ATM con- nection depend on the connection corresponding traffic parameters. “Worst case” traffics are generally used to provide robust procedures [3]. In this work, we assume that CAC on a multiplexing ATM node is based on “worst case” traffic represented by a set of On/Off sources connections and that QoS indicator is the cell loss probability. The multiplexing node is then presented like a queue fed by bursts of cells. Many authors model this multiplexing scheme using fluid models at burst level [l]. Generally, homogeneous On/Off sources are considered. These models analyse the be- haviour of a queue fed by a finite number of identical On/Off sources where traffic pa- rameters are the Peak Cell Rate, PCR, and the Sustainable Cell Rate, SCR. We notice on the results derived from these models an asymptotic behaviour when we display the admissible load versus sources activity rate for a given loss probability. An asymptote ap- pears when the activity rate of a single source, SCR/PCR, becomes very small. We take advantage of this limit to construct a simple batch model approaching the multiplexing of a large number of On/Off sources with very low activity rate (i.e. SCR N 0). We propose a model which is a finite queue fed by bursts of cells. In order to model batch arrivals of an infinity of sources with very low activity rate, we consider a Poisson arrival process. The simple model M/D[k]/l/N has first been developed [2]. In this model, we assumed that each customer represented a burst. This modelled the multi- plexing of homogeneous bursts. We extend now this model to the case of heterogeneous sources and we propose the cell level model M[lcl] + M[lcz]/D[lc]/l/N where two sources 1