TIME SERIES MODELS: FORECASTING AND CONTROL FOR THE MANAGEMENT OF COMPUTING INFRASTRUCTURE AND RESOURCES

Abstract

The existence of self-similar or fractal nature of network traffic has been proven by recent studies over a wide range of time scale. These properties are very different from the traditional Poisson processes/models. If one analyzes the network traffic by these traditional Poisson models then this leads toward wrong or incorrect decisions. The issues with Poisson models are overestimation of the performance of computer network and false allocation of data processing and network resources. Hence complete understanding of self-similar nature is required for fast and appropriate performance of the network. This thesis describes a number of problems arising from the experimental study of telecommunication networks. The work can be split into two main areas: full understanding of self-similarity within a network traffic and secondly focused over the comparison of two different networks i.e. searching similarity between two or more self- similar network traffic by using wavelet based time warping technique. Estimation techniques of self-similarity measure i.e. Hurst Index H is considered first the critical analysis of these methods is studied. There is a lack of GUI based tool for the estimation of Hurst Parameter we developed a new Statistical tool that can estimate the Hurst Parameter H by different methods along with their statistical analysis. Next we focused over the generation of self-similar traffic using simulation methods. A comparison of these methods was also studied. Main focus was on the Wavelet based estimation and generation methods for computer network traffic. Secondly concentration is made over filtering of temporal data to get the smooth form of the signal that may be used for further analysis. The main goal of second part is to explore new trends for filtering and smoothing of temporal data. A recently developed technique of wavelets for smoothing of temporal data is explored. New approach was developed using different features of wavelet transformation for dimensionality reduction and tested them for various decision making processes in business and computer science. The developed smoothing techniques are applied for forecasting, query processing of vsimilar sequences in huge databases and their clustering on the basis of membership value. The developed forecasting technique is based on Wavelet families and Seasonal Autoregressive integrated moving average model. In this thesis, two new approaches for smoothing of large temporal databases using the wavelet filtering theory for approximations of query procedures for decision support systems (DSS) are also proposed. For these smooth signals, a novel query processing algorithms was developed while selecting wavelet based features and time warping distance metric, called wavelet based features time warping. For first proposed algorithm we utilized the features extracted on the basis of minima, maxima and averages of wavelet based compressed signals and for second proposed algorithm local features of wavelet transformation using average of approximation coefficients at the coarsest scale and maxima of maxima and minima of minima of detail coefficients at all scales were used. Our both models support index based time warping distance. It is proved by carrying out extensive experiments with synthesized databases using different wavelet families that the proposed methods are very effective and ensure the nonoccurrence of false dismissals and minimal false alarms with least compromise over accuracy. The developed method gives extremely fast response times as our approximate query executes its maximum processing over synoptic set of wavelet features. A new measure of degree of similarity is introduced for clustering using time warping distance method which gives better control over the number of clusters.