Research On QoS-aware Traffic Classification And Aggregation

Traffic classification is necessary for effective network monitoring,resource management and security control.It has become a research focus in the field of network communications.The conventional traffic classification technologies have become nearly obsolete due to the limitations in processing encrypted flows and user privacy.Most recent studies have focused on flow-feature-based methods.According to different extraction pattern of flow features,they can be further divided into statistical features based methods and deep learning based methods.However,with the increase of the number of classes,statistical features based methods fall into trouble because of "feature engineering";while for deep learning based methods,the target classes should be fixed,and the training process is extremely complex.In addition,deep learning does not have the ability of self-learning and self-adaptive,which is not suitable for online network classification.Therefore,this thesis deeply analyzes the ground truth of network traffic,and explores the fundamental reasons of the dilemma in traffic classification,and develops elegant solutions to the problems on the basis of data correlation.Thus the fractal theory is introduced to achieve an innovative improvement of traffic classification.In addition,the main purpose of traffic classification is to implement differentiated services,and enforce end to end QoS in heterogeneous network environment.The results of transmission without considering QoS are that: flows in the same aggregate may have different QoS requirements;while some flows in different aggregates may have the same QoS requirements.Such results are not conducive to the execution of end-to-end QoS support,and also turn out to be contrary to the QoS framework.Therefore,transmission of flows would be driven by QoS.The existing flows mapping methods generally exploit quantitative means,and they require explicit values and weights for QoS parameters.However,these factors are typically uncertain and imprecise in reality.Consequently,a qualitative flow mapping method based on the preference logic is proposed in this thesis to address the above issues.The main research work and contributions are as follows.1)The flow fractal is proposed with rigorous theoretical proof.A large number of observations and analysis of traffic flows show that the failure of statistical features lies in the independence hypothesis.That is,the continuously arriving packets are supposed to be statistically independent,the size of packets and the interval time of packets are also supposed to be independent.However,there is evidence of a strong correlation between the data.Therefore,based on the packets correlation,the flow fractal is established to subvert the independence hypothesis,which provides theoretical basis for flow fractal characteristics being used in traffic classification.2)Traffic classification based on the fractal exponent.Many studies show that the type of bitstream is affected by the transport layer protocol,congestion control mechanism,and retransmission of lost packets,which uausally cause a specific flow shape for a certain type of bitstream.Fractal exponent is a general representation of fractal things,which is used to describe the degree of self similarity.The fractal exponent is used to roughly reflect the overall difference between the flow shapes and thus helps to achieve classification of flows.The experiment results show the classification accuracy of the proposed method reaches high than 97% when classifying four types of video flows,including QQ,PPlive,GAME,and BT.Therefore,the fractal exponent has practical significance for coarse classification.3)Traffic classification based on the fractal exponent in the wavelet domain.Note that the fractal exponent can roughly reflect the overall difference between flow shapes,and thus can be used for coarse classification.However,with the increase of the number of classes,the difference between flow shapes decreases,and then the classification becomes unstable due to the fuzzy fractal exponent.Therefore,a precise computing model is proposed to obtain accurate fractal exponent in wavelet domain to improve the classification performance.In this model,the estimated fractal exponent in wavelet domain is derived,the optimum segments based on cost function is analyzed,and the statistical differential level between the segmented fractal exponents is calculated with the method of clustering.Finally,the classification is implementd with maximum between-cluster variance threshold.The result shows that the classification method with the fractal exponent in the wavelet domain is comparatively stable and has strong adaptability when the experimental environment changes.4)QoS-aware fine classification based on the fractal spectrum.The fractal exponent is a macroscopic representation of flow shape,which cannot be applied to fine-grained network traffic classification due to its lack of detailed descriptions.Therefore,the fractal spectrum is introduced to describe the complex fractal characteristics of flows,which represents the detail characteristics on bursting data,and thus to achieve fine classification.However,it is difficult to calculate the fractal spectrum of flows.In this thesis,the estimated spectrum is obtained by numerical analysis: First,scaling function ?(q)of flow is defined;Then the relationship between fractal spectrum and ?(q)is deduced based on classical Legendre transformation;Finally,?(q),which are comparatively easy to calculate,is expected to describe the complex fractal characteristics of flows and thus implement classification at the fine grained level.In the dynamic performance testing phase,500 flows are selected randomly from 20 QoS classes,such as instant video and streaming media.The results show that fractal spectrum can achieve a superior performance for QoS-aware fine classification.5)QoS-aware soft classification in granular domain.In dynamic networks,problems such as loss,retransmission,and disorder of packets may occur at any time,which causes the testing data has a big deviation against the training samples,and thus cannot be correctly identified.Therefore,granular computing is introduced and thus achieves an improved soft traffic classification model.In this soft model,the granules of traffic are defined,and then the fractal characteristics among the granules are analyzed to achieve the granular relation matrix(GRM).The traditional statistical features turn out to be a special case of GRM at the maximum scale.GRM describes the flows more comprehensively and thus classifies them more accurately.Compared with baseline methods,including I-SVM,K-L,SFNN,etc.,the experimental results demonstrate its superiority for dynamic QoS classes.6)QoS-aware flow aggregation method using preference logic.Transmission of aggregated traffic must be based on QoS.In this thesis,a qualitative method based on the preference logic and QoE,is proposed to model the QoS requirement of the network flows.Then the most suitable QoS queues are reasoned by non-monotonic among the candidate QoS queues,and thus the dynamic flows aggregation is realized in a qualitative way.The experiment results show that the proposed method can effectively model the QoS requirement of the flows,and adjust the aggregation process to make the best use of the limited system resources under a highly dynamic environment,where the value and weight of the QoS parameters are changeable,and the QoS queues are varied.Compared with the existing methods of flows aggregation,the proposed preference logic based aggregation scheme performs better in terms of delay,packet loss rate and throughput.