Abstract: More and more researches show that network traffic is not Markovian process, but shows the burst nature called self-similarity at any time scale .The mathematic models describing network traffic mainly include self-similar process and long range dependence structure. Due to some network parameters obeying heavy tail distribution, network traffic shows the burst nature at large time scale. This paper analyses the drawbacks of the classical network traffic models and performance evaluation, and describes the basic trademarks of evolutionary traffic models. This paper studies three important models of self-similar network traffic: ON/OFF model, user access probability model and fluid flow close loop model, and discusses relative research directions. Some issues and principles that shall be noticed during studying and modeling network traffic are given in the end.