In this paper, we study the performance of packet-level scheduling in IP-over-ATM networks with QoS control. That is, we assume the traffic is composed of multiple classes. We analyze the performance of different queue mappings between traffic types and the number of available traffic classes (priority queues). Since cells of a given packet are not bound to be transmitted back-to-back if multiple traffic classes are used, it is quite interesting to know the packet delay characteristic as well as the cell delay characteristic of respective traffic types in different queue mappings. Closed-form solutions for the mean cell waiting time and the mean packet waiting time of individual traffic types in different queue mappings are presented. The numerical results obtained in this paper can be helpful in understanding the behavior of IP-over-ATM networks which adopting packet-level scheduling and QoS control.

We propose a finite-capacity single-vacation model, with close-down/setup times and a Markovian arrival process (MAP), for SVC-based IP-over-ATM networks. This model considers the SVC processing overhead and the bursty nature of IP packet arrivals. Specifically, the setup time corresponds to the SVC setup time and the vacation time corresponds to the SVC release time, while the close-down time corresponds to the SVC timeout. The MAP is a versatile point process by which typical bursty arrival processes like the IPP (interrupted Poisson process) or the MMPP (Markov modulated Poisson process) is treated as a special case. The approach we take here is the supplementary variable technique. Compared with the embedded Markov chain approach, it is more straightforward to obtain the steady-state probabilities at an arbitrary instant and the practical performance measures such as packet loss probability, packet delay time, and SVC setup rate. For the purpose of optimal design of the SVC-based IP-over-ATM networks, we also propose and derive a new performance measure called the SVC utilization ratio. Numerical results show the sensitivity of these performance measures to the SVC timeout period as well as to the burstiness of the input process.