An effective QoS-constrained IP network must incorporate the time-priority scheduling paradigm. Policies based on priorities can be exerted to different layers of a packet switching network. For example, at the admission control layer, real-time applications should have the priority higher than non-real-time ones to get the required connection. Preemption is associated with priority. A scheduling discipline is non-preemptive if, once a stream has been given the service like a transmitter, the service cannot be taken away until the job is complete. It has been well known that, at the layers other than the connection layer, the traffic usually presents a self-similar (SSM) behavior. One primary attribute of the SSM traffic is the heavy tailed (HT) distributions. In this paper, we propose several capacity allocation models taking the following features into account: (1) The packet inter-arrival time follows the exponential distribution; (2) The packet length follows the Pareto distribution; (3) There are multiple priority classes; and (4) The low priority class can be preempted by a high priority class. The new models are mainly used at the connection level due to feature (1). However, the new models distinguish themselves from those conventional ones due to features (2), (3), and (4).