On network resource allocation using alpha-stable long-range dependent traffic models

Abstract

Recent studies suggest that networks should be designed taking into account the long-range dependence and high-variability properties of the traffic they carry. It has been proven in the past that these two statistical properties can be properly represented using traffic models based on alpha-stable self-similar stochastic processes. Assuming this traffic modeling approach, in this dissertation we propose and evaluate some techniques for resource allocation. We propose suitable envelope processes for the levels of bandwidth demand which allows us to develop static resource allocation schemes. The proposal is based on a generalization, to the alpha-stable case, of the concept of probabilistic envelope processes, which have been previously defined for simpler models. It is shown that, with this approach, we can simply and effectively deal with much of the argued complexity encountered in alpha-stable models and develop techniques for proper dimensioning of network elements. From our analysis it is concluded that the presence of heavy tails in the distribution of the traffic process has a severe impact on the requirements of network resource. For instance, the multiplexing gain is negatively affected, which directly impacts the scale economies expected by service providers. In order to cope with these issues, dynamic resource allocation is also considered. A dynamic prediction-based resource allocation method is introduced and evaluated. It is shown that it significantly improves network utilization over static resource allocation schemes in trade for some signaling and processing overhead. Although other schemes based on prediction have been proposed, we use a novel linear prediction algorithm for symmetric fractional stable noise. This approach is intended for some traffic classes whose marginal distribution exhibits a heavy tail. The linear prediction algorithm we use was recently introduced by other researchers, but has not been studied in detail. Therefore, its performance evaluation is also carried out. In addition to this study on the prediction-based approach, a dynamic resource allocation scheme based on envelope processes is also introduced and evaluated. We conclude that when alpha-stable models are properly used and interpreted, they let us accurately represent network traffic and therefore design and analyze reliable resource allocation mechanisms.