A Semi-Markov Model For Network Traffic

As a basic research in networking area, traffic modeling is fundamental for networkperformance analysis and planning. A fine traffic model is very important to planhigh-performance protocol, to design an efficient topology, to provide high quality ofservice with high ratio of performance and cost, to accurately analysis performance andpredict traffic, and to control congestion and smooth traffic. Traffic modeling is also animportant component of the modern network management system, which can make usunderstand network characteristics, accommodate network resource rationally and realizeefficient network management. The core of traffic modeling is getting key stochasticproperty in network traffic. Although computer network traffic modeling and queueperformance have been studied for more than ten years, it has not got a consistentconclusion in network traffic modeling. None of traffic models can be widely approvedand applied in IP networking area, just like Poisson Model in telecommunications does.The thesis presents a new traffic model based on semi-Markov Process to describenetwork traffic characteristics, and catch traffic property accurately. It is also tractablein computing. Based on the model, we could understand traffic behavior better, predicttraffic more accurately, manage and allocate network resource more effectively.Innovations of the thesis are as follows.1. A Semi-Markov Model for network traffic is presented. According to trafficcomponents and characteristics in different stages, the model divides network trafficinto four states: busy, idle, rising and falling, through setting busy and idle thresholdsto network traffic. In such way, key stochastic properties are more distinctive andeasier to be induced. Traffic characteristics in different states and relationshipsbetween states are explored to describe the network traffic Semi-Markov modelcompletely. According to the influence of protocol on traffic in each state,especially on change of traffic rate, the following is assumed: the average traffictransmission rate under a busy state follows geometric brown motion, that under anidle state it follows a normal distribution, and that under rising or falling states itfollows exponential distributions respectively. To verify above assumptions and todeduce methods of estimating parameters of the model and their physical meanings,some international general data are analyzed. The statistical analysis shows thatabove 95% traffic data has the characteristic that hypothesized under itscorresponding state, and the hypotheses are acceptable.2. A very important network performance criterion, network utilization, is computedbased on the model. As an application of the model, the paper describes how tocalculate network utilization, the ratio of current network workload to maximaltheoretic workload, based the Semi-Markov Model for network traffic. As weknow, network utilization estimation is a very important network performancecriterion The experimental results based on analysis and validation of threeinternational general data show that the relative error between realistic networkutilization and theoretical one calculated by the model is less than 5%, whichindicates how well the Semi-Markov Model describes network traffic characteristic3. As another application, the model is applied to tranffic prediction to validate themodel's feasibility and correctness. Traffic prediction is widely used in networkapplication and network management, and is very important to network performanceanalysis and network control. This thesis emploies the Semi-Markov Model topredict network traffic. The traffic prediction algorithm based on the Semi-MarkovModel focuses on predicting the possible upper bound of traffic rate rather than anaccurate value at some future moment. In fact, it is difficult and unnecessary topredict an accurate value for the future traffic, an upper bound should be goodenough. Based on this idea, an upper bound of traffic rate prediction formula isgiven and verified by actual trace data. The experimental results show that predictionaccuracy reaches 80%, especially that of short-term reaches 90%, which denotes thatour method can be suitable for both short and long term prediction. Deviation degreeresults show that most of relative errors are small, especially for backbone networks,the relative errors can be lower than 15%.It is very difficult to model network traffic due to its heterogeneousness andburstiness. The research of the Semi-Markov Model has just begun and many worksshould be carried on. Especially more practical applications based the model should bedeveloped. Even so, as a basic research in networking area, traffic model is veryimportant and has good prospects to understand network behavior, to improve networkperformance and to plan high-quality network.