G. FODOR ET AL.: ON PROVIDING BLOCKING PROBABILITY AND THROUGHPUT GUARANTEE IN A MULTI SERVICE ENVIRONMENT 1 On Providing Blocking Probability- and Throughput Guarantees in a Multi-service Environment G´ abor Fodor , S´ andor R´ acz , Mikl´ os Telek Ericsson Research, Sweden, Gabor.Fodor@era-t.ericsson.se Technical University of Budapest, Hungary, raczs@ttt-atm.ttt.bme.hu, telek@hit.bme.hu Abstract As the Internet evolves from a packet network supporting a single best effort service class towards an integrated infrastructure supporting several service classes - some with QoS guarantees - there is a growing interest in the introduction of admission control and in devising bandwidth sharing strategies, which meet the diverse needs of QoS-assured and elastic services. In this paper we show that the extension of the classical multi-rate loss model is possible in a way that makes it useful in the performance analysis of a future admission control based Internet that supports traffic with peak rate guarantee as well as elastic traffic. After introducing the model, it is applied for the analysis of a single link, where it sheds light on the trade-off between blocking probability and throughput. For the investigation of this trade-off, we introduce the throughput-threshold constraint, which bounds the probability that the throughput of a traffic flow drops below a predefined threshold. Finally, we use the model to determine the optimal parameter set of the popular partial overlap link allocation policy: we propose a computationally efficient algorithm that provides blocking probability- and throughput guarantees. We conclude that the model and the numerical results provide important insights in traffic engineering in the Internet. Keywords Bandwidth sharing objectives, multi-rate loss models, blocking probabilities, throughput, Markov reward models. I. I NTRODUCTION In recent years there have been significant advances in researching and standardizing mechanisms that are capable of providing service differentiation in the Internet. While there still seems to be a wide span of the methods which aim at providing QoS differentiation between contending flows, it is widely accepted that there is a need for traffic engineering mechanisms which control the access of the different traffic classes to network bandwidth resources. In particular, there is a growing interest in devising bandwidth sharing algorithms which can cope with a high utilization in the network and at the same time take into account the different traffic classes’ throughput and blocking probability requirements. Recent research results indicate that it is meaningful to exercise call admission control (CAC) even for elastic traffic, because CAC algorithms (and consequently the blocking of some arriving flows) provide a means to prevent e.g. TCP sessions from excessive throughput degradation [16], [17]. From this perspective it is important to develop models and computational techniques that make analytical studies of the behavior of such future types of networks possible. Generally the issue of bandwidth sharing should be considered in the context of dynamically arriving and departing flows, which naturally calls for the application of the classical multi-rate loss models. These models have proved useful in the dimensioning and performance evaluation of circuit switched as well as ATM networks. Thus, they provide motivation for extending the applicability of this modeling paradigm to Internet context. Unfortunately, a direct application of the multi-rate models for traffic engineering in the Internet is non-trivial, because: By definition, it is not possible to associate a constant bandwidth with elastic services, like the best effort (without minimum rate guarantee) or the ”better than best effort” (with minimum rate guarantee) type of services. The bandwidth occupied by the elastic flows depend on the current load on the link and on the scheduling and rate control algorithms applied in the network nodes. The notion of blocking, when applied to elastic flows, needs to be reconsidered because an arriving elastic flow might get into service even if at the arrival instant there is no (or very small) bandwidth available. For many services, we need to take account of the fact that the actual residency time of the elastic flows depend on the throughput which the flow receives. For instance, an ftp session would last longer if its throughput decreases. (Real-time M. Telek was partially supported by OTKA T-34972.