| Recent experimental studies reveal that traffic in modern high-speed network exhibit self-similarity. The burstiness of these traffic exists over a wide range of time-scales. The long-range dependence they possess renders conventional Markovian models, which are short-range dependent in nature, inadequate. In particular, it is found that traditional models seem to overestimate system performance in case the traffic is self-similar. This sparks a new area for researchers to model, analyze, and control self-similar traffic. Although self-similar traffic models are parsimonious, they are in general more complicated to analyze than Markovian models, which have the nice memoryless properties. Hence, there are relatively few analytical results in the literature for fractal queues, and most of these work are confined to a single queue. Inspired by the robustness of self-similarity (merging and splitting of self-similar processes are self-similar), we study the decomposition of a network loaded with self-similar traffic. Through simulations, we demonstrate that an estimation for the end-to-end delay can be obtained through decomposition. The decomposition method provides a practical tool for traffic engineers to analyze high-speed networks with realistic self-similar traffic, which is otherwise not tractable.;Another major implication of self-similar traffic is the existence of correlation over time. This makes possible short term prediction on the future traffic load. We study a measurement-based congestion alarm for self-similar traffic. Like ordinary measurement-based admission control mechanisms, we perform traffic measurement to determine the actual traffic load regularly. Modeling the system aggregate traffic by a self-similar process, short term prediction can be performed. So when the system capacity is unlikely to handle the anticipated traffic loading in the near future, the alarm will be set off and appropriate network management functions need to be carried out in order to alleviate the situation. In our study, we consider the use of dynamic bandwidth allocation whereby the system bandwidth is renegotiated. Our simulation results show that our proposed scheme requires about 20% less capacity on the average than a system with fixed capacity under different scenarios. |
