The Study Of The Self-Similar Network Traffic: Characteristics And Its Impact

Since the seminal study paper of self-similar phenomenon in network traffic waspublished in 1994,many researchers' attention had been paid to explain howself-similarity traffic occurred.Some detection methods had been proposed.How theself-similar traffic affected the network performance,including Internet,Optical BurstSwitching (OBS) network,and wireless network,had been studied.Especially thequeue performance (packet loss rate,delay,and delay vibration) and relevant algorithmshad been analyzed deeply under the self-similar traffic.Many theories and methods forimproving queue performance under the self-similar traffic had been proposed,andsome simulations had been reported.In this thesis,starting with the diverse definitions of self-similarity,we try to find outtheir relationships in order to lay a foundation for studying the self-similar trafficfurther.The main results are as follows:1.The distribution of file sizes is not always close to the Pareto distribution (heavytail),it is closer to the lognormal distribution (not heavy tail).Not only the heavy taildistribution of file sizes can produce self-similarity,but also the non-heavy taildistribution may result in self-similarity.2.Several physical models and stochastic models,which can be used to explain theself-similar phenomenon,have been summarized.The complexion of obtaining theparameters of these models has been evaluated,which is the basis of measuringself-similar traffic in networks.3.The assembly algorithms with fixed time threshold,the assembly algorithm withfixed length threshold,the united algorithm and assembly algorithm with adaptivelength threshold in OBS network have been summarized.A modified assemblyalgorithms with an adaptive length threshold has been put forward,which can reducethe average packet block rate about 10 times more than other algorithms under theself-similar traffic.4.The change of Hurst parameter of self-similar traffic has been analyzed when the network gateway works on energy conserving model or instant transfer model under theself-similar traffic.The theory analysis and simulations show that"the trafficself-similarity does not change as it enters the wireless network when the gatewayworks on the certain state"is not correct,which A.P.Petropulu provided.The morereasonable result is given.5.Two prediction methods,which are derived from the FARIMA models and theneural network based on the self-similar traffic,has been analyzed,a predictionstructure and the weight calculation method are proposed based on five layers' neuralnetwork.The theory analysis and simulations show that when it is used to managebuffer appropriately under self-similar traffic,the packet loss can be reduced,and thefairness can be maintained.6.The effective bandwidth estimation method under the self-similar traffic isderived out.The simulations show that the bandwidth estimation method based onlognormal-FARIMA model can satisfy the overflow rate limitation.The affection of self-similar traffic to network performance,especially queueperformance (such as buffer size,delay,packet loss rate) and algorithms (such asRandom Early Detection),some important transfer system (such as Transport ControlProtocol/Internet Protocol),and resources management (such as bandwidthmanagement),will be worth to study further.