Measurement-based Traffic Characterization and Resource Allocation

点击翻译按钮获取中文摘要

Compared to traditional voice and data networks, next generation networks will present more difficult challenges to traffic control because (1) the carried traffic will be much more diverse and complex in nature, (2) traffic rates will be faster, and (3) network performance levels must be protected for certain applications requiring a quality of service (QoS). Traffic analysts attempt to gain an understanding of the stochastic nature of the traffic from volumes of collected traffic measurements, and derive traffic models and control algorithms to manage network resources (such as connection admission control in ATM networks). Unfortunately, traffic analysis is typically a time consuming process and requires a high level of statistical expertise. Many stochastic traffic models have been studied but methods for evaluating their validity for real traffic are generally lacking. The objective of the proposed research is to improve understanding of real network traffic by developing systematic methods for processing raw traffic measurements, deriving traffic characteristics, and applying these characteristics to real-time resource management. The approach is based on a modular library of traffic models and statistical estimation techniques to fit and test eachmodel to observed traffic. The proposed research project will be carried out in three phases. The first phase will develop a modular library of traffic models including the well-known Markov on-off model, Markov modulated Poisson process (MMPP), Markov modulated fluid, autoregressive process, and fractional brownian motion or other self-similar models. Each model consists of a number of estimatable parameters and a set of structural properties. Research in the first phase will derive parameter estimation methods for each traffic model using maximum likelihood estimation or Bayesian estimation, and evaluate the validity of each model using hypothesis testing, chi-square goodness of fit tests, and other tests depending on the particular properties of each model. Given raw traffic measurements, the best-fit traffic model in the library can be methodically selected at any time. The second phase of the proposed project will derive efficient sequential estimators from the first phase and implement these estimators in software for real-time traffic analysis and visualization. The software will work with a hardware traffic collection system to infer general statistical characteristics (such as burstiness) and display the best fitting traffic model. The traffic model will reveal the underlying behavioral properties of the traffic. Finally, the third phase will study real-time resource allocation using the estimation methods and software from the earlier phases. Instead of assuming a traffic model a priori for resource allocation, the proposed method will dynamically select the best fitting traffic model from a library and adapt resource allocation decisions in real time. This approach may eliminate the uncertainty and inaccuracy in the traditional static approach. The method will be studied by means of OPNET simulations for the specific problem of ATM connection admission control. Successful results from the proposed project will be software tools and algorithms to help networkadministrators gain insight into the nature of network traffic and fine-tune their control of network resources. The project will also be useful to researchers, educators, and students by providing tools to examine real traffic and understand underlying behavior.