1 Abstract—Quality of service architectures that provide efficient resource allocation for multiple services have been an active research topic for a long time, but unsuccessful when it comes to deployment. The goal in this paper is to propose an admission control algorithm that both preserves the quality of service and efficiently shares resources based on the demand in different classes. Based on the concept of externality costs we derive a solution to balance the admission control in a high priority class with congestion control for best effort traffic. A low complexity implementation based on measuring the load of high priority traffic and the loss in the best effort class is proposed. It is then evaluated by simulations which show that it can achieve an efficient sharing between the traffic classes. Index Terms—Admission control, congestion control, network pricing, quality of service. I. INTRODUCTION ALL admission control is essential to satisfy the quality requirements of multimedia applications that run over packet switched networks. Most admission control solutions have as goal to provide good quality to admitted traffic by keeping the load of the network below a predefined level. For distributed admission control it is better to characterize the state of each network element in a way that can be aggregated to reflect the state of the whole path. The characterization can be defined based on concepts from economy or optimization theory that describes how to maximize the total utility of the network, as described by Kelly [14]. In this paper externality costs are used to define an admission control algorithm. Externality costs reflect the degradation, like packet delay, packet loss or throughput reduction, that one user cause to the other users. The externality cost caused by the high priority traffic on the best effort class is shown to be proportional to the congestion signal in the best effort class. Therefore, the algorithm turns out to have a practical measurement based implementation. The proposed solution is considered in the context of the currently ongoing work on admission control for Diffserv [1]. In the IETF the proposed distributed solution is based on pre- congestion notification (PCN) [5]. In PCN the core routers of H. Lundqvist is with NEC Laboratories Europe (e-mail: henrikl@kth.se). The research results presented herein have received support from Trilogy (http://www.trilogy-project.org ), a research project (ICT-216372) partially funded by the European Community under its Seventh Framework Programme. The views expressed here are those of the author(s) only. The European Commission is not liable for any use that may be made of the information in this document. a domain are flow unaware and implements functions that mark packets in a traffic class. The favored marking methods are based on token bucket or virtual queue implementation in the routers. In the initial PCN draft [2] it was also proposed that the best effort load could be used in the PCN marking function. However, in the later versions that feature has been dropped in favor of lower complexity. In a longer perspective it is valuable to take into account also the best effort traffic since it can be used to provide an automatic resource reallocation between the classes. The solution is also applicable for flow aware admission control, for example in access networks. In this case an additional benefit is that the consequence of admitting the new flow on the admission metric can be estimated in advance by using a limited flow description, such as the peak rate. The main contributions of this paper are: 1. It is shown that the best effort loss rate is an appropriate measure for the externality cost caused by high priority traffic. 2. A practical admission control solution is proposed based on effective bandwidth and time scale composition concepts which are not considered in current PCN standardization efforts. 3. The sensitivity to certain parameters is studied through simulations and the adaptive resource allocation is compared to the currently proposed solution. In the next section the main application scenarios are described to provide the background for the solution. Section III describes pricing and resource allocation and explains how the best effort congestion level should be taken into account in the admission control. In Section IV the control and timing issues are treated and a design of the admission control algorithm is presented. Then a simulation study demonstrates the properties of the solution for both flow unaware and flow aware networks. Finally the paper is concluded in Section VI. II. BACKGROUND AND APPLICATION SCENARIOS The proposed solution is intended for multi class networks, typically using Diffserv [1]. Since Diffserv can be used in various settings in different parts of the network the following section describes the terminology used in the paper and how the solution can be implemented in different scenarios. A. Terminology We use the term decision point for the entity where the admission control decision is taken. The decision can be made locally in each node and then be signaled to the end-points, or distributed for a path through multiple nodes. Here we Admission Control with Resource Reallocation Henrik Lundqvist NEC Laboratories Europe C