to 
depending on the kind of application).
The integration of heterogeneous applications within a single network
requires a complete rethinking in network engineering and traffic management.
Indeed, traditional computer networks, such as the Internet, have been
designed for the transportation of a single kind of data traffic, also
called best effort traffic. The number of active sessions in such networks
is unlimited and no kind of priority policy is implemented. However, modern
high-speed networks should be able to support different degrees of Quality
of Service (QoS) to different applications. For example, real-time traffic
generated by multimedia applications has radically different requirements
than best-effort traffic [Kar96]. First,
real-time applications require tight bounds on transfer delay (in the order
of hundreds of milliseconds). Second, the loss probability of network packets
belonging to multimedia applications must be very small (varying from
to
depending on the kind of application).
The needs for QoS guarantees in high-speed networks is not limited to real-time multimedia applications. For instance, many users are willing to pay a larger amount of money in order to get high quality data services guaranteeing both high throughput and low delay. The emergence of this new market is a consequence of the explosion of World Wide Web applications.
In order to overcome the current limitations of the Internet architecture, an Integrated Service (IntServ) model has been recently standardized (RFC's 2210-2215) for supporting real-time services. The IntServ model includes two new sort of services targeted towards real-time traffic: guaranteed and controlled-load services. The guaranteed service [SPG97] is intended to provide strict (deterministic) guarantees on QoS requirements while the controlled-load service [Wro97] will provide only rough statistical guarantees. The main advantage of the controlled-load service over the guaranteed service is that it can achieve a higher network utilization. Both guaranteed and controlled-load services are based on a ``Resource reSerVation Protocol'' (RSVP) (see [Zha93] and RFC's 2205-2210) and on a ``Call Admission Control'' (CAC) procedure (see [PeE96] for a review of some CAC schemes). Before admitting any new session, the RSVP protocol checks if the network has enough free resources to allocate to the new call and the CAC algorithm makes sure that QoS requirements of on-going calls will not be violated after the new call establishment. The implementation of the IntServ model seems to encounter many difficulties due to the important changes it requires from the current Internet service model, especially with regards to call set-up and router architectures. An intermediate solution, referred as Differentiated Service (DiffServ) [KLS98], is currently under the way of standardization.
The ITU has chosen an alternative network architecture called ATM (Asynchronous Transfer Mode) [Pry93, ATM96] as the target transfer mode approach for providing the integration of the various traffic types to be supported by B-ISDN. ATM is a connection-oriented packet-switched mode of transfer using 53 bytes length packets called cells. All the cells belonging to the same connection follow the same path (virtual circuits) along the network. ATM is, therefore, naturally suited for implementing resource allocation algorithms and call admission control. Moreover, the ATM forum has recommended a set of four layer service classes to cover the QoS service requirements of the various types of traffic expected to utilize ATM networks facilities. Two of them are mainly intended for data traffic (ABR and UBR). The CBR service layer is designed for traffic, such as uncompressed voice or video, with Constant Bit Rate and with tight constraints on end-to-end transfer delay and loss probability. Finally, the VBR service layer is designed to accommodate the important class of Variable Bit Rate traffic having the same type of constraints as CBR traffic. One of the key example of VBR application is a compressed video application like MPEG.
A salient feature of VBR traffic is that its peak rate can be substantially higher that its average rate. A call admission control based only on the peak rate of each session would be very conservative since most of the bandwidth reserved by the session would be wasted. A way to increase network utilization is to multiplex VBR sources. However, multiplexing and sharing of common resources by different network sessions imply possible degradation of QoS perceived by each individual session. The problem can be thus described as how to maximize the utilization of network resources while preserving the Quality of Service guarantees of each individual session.
In order to answer effectively the question of maximizing network utilization, one has first to employ accurate models of traffic streams arising in high speed networks. Second, based on the assumptions suggested by these models, one has to develop advanced mathematical tools for the efficient estimation of network performances measures related to QoS parameters.
