Efficient Resource Management for Hard Real-Time Communication over Differentiated Services Architectures Byung Kyu Choi Riccardo Bettati Department of Computer Science Texas A&M University College Station, TX 77843-3112, USA choib, bettati @cs.tamu.edu Tel:1-979-845-5469, Fax:1-979-847-8578 Abstract We propose an efficient strategy for resource manage- ment for scalable QoS guaranteed real-time communication services. This strategy is based on sink trees, and is partic- ularly well suited for differentiated-services based architec- tures. We first show that finding a set of sink-trees in a given network is NP-complete. Then we propose a heuristic algo- rithm that always efficiently produces a set of sink-trees for a given network. Sink-tree based resource management in- tegrates routing and resource reservation along the routes, and therefore has a number of advantages over other re- source management scheme, in terms of: admission proba- bility, link resource utilization, flow set up latency, signaling overhead, and routing overhead. In this paper we show by simulation experiments that even for simple cases the sink- tree based approach shows excellent results in terms of ad- mission probability. 1. Introduction This paper addresses an efficient resource management over computer networks in which scalable QoS guaranteed real-time communication services are provided. In this con- text, by real-time we mean that packets delivered beyond their end-to-end deadline are considered useless. Internet- working technology is increasingly being used for applica- tions with this kind of requirements, for example, Voice over IP, military communications, and industrial control systems. Traditionally, best-effort service has been the main type of service available over internetworks. While this type of service has contributed much towards the rapid growth of the Internet, it can-not support applications that have real- time requirements. A systematic approach based on the connection admis- sion control and packet scheduling was proposed within the IntServ architecture of the IETF [3] to accommodate the real-time requirements. However, the lack of scalability in IntServ is the cause for its very limited deployment to date. This limitation is being addressed within the DS (Differenti- ated Services) architecture [8, 11, 12]. From the user’s point of view, the DS model partitions each user-level flow into one of a set of predefined classes. Packets of each class are served inside the network according to a class-based scheduling policy. The result is that routers are aware only of aggregations of flows. Each edge router is supposed to aggregate the individual flows into a small number of such aggregate flows. In this fashion, the DS model makes the network scalable regardless of the number of flows. In order to provide connections with bandwidth and de- lay guarantees, an admission control mechanism must be in place. The admission control procedure can be centralized or distributed. However, in order to be effective, it must be aware of every node’s current workload and admissible resources [13]. The more such information is available to the admission control procedure, the higher is the letter’s effectiveness in terms of admission probability of new connections. There is an obvious trade-off between the effectiveness of such a pro- cedure and its efficiency, as collecting and maintaining this information would require significant signaling overhead. There is a research work reporting that signaling and resource management should be taken into account at the same time to maximize the resource utilization in a network [14]. Any admission control procedure inevitably includes some amount of signaling overhead between the decision In the following, we will use the term flow to indicate a stream of data between a source and a destination, and the term connection to indicate the virtual circuit that needs to be established to carry the flow.