Ethernet traffic is statistically self-similar, meaning it exhibits similar patterns across all time scales. This behavior is not captured by conventional traffic models and has significant implications for network design, control, and analysis. Self-similar traffic lacks a natural "burst" length, with bursts consisting of bursty subperiods separated by less bursty subperiods at every time scale. Analysis of hundreds of millions of Ethernet traffic measurements from 1989 to 1992 shows that self-similarity is evident, with the Hurst parameter indicating the degree of burstiness. Self-similar traffic is different from conventional telephone traffic and standard packet traffic models, such as Poisson or Markov-modulated Poisson processes. The burstiness of LAN traffic typically increases with more active sources, contrary to common assumptions. Self-similar models provide accurate and realistic descriptions of traffic scenarios in high-bandwidth networks. The paper discusses the implications of self-similar traffic for congestion control, showing that congestion management in self-similar environments differs from conventional models. The paper also presents statistical methods for analyzing self-similar data, including R/S analysis, variance-time plots, and periodogram-based methods. The results show that Ethernet traffic is self-similar with Hurst parameters ranging from 0.75 to 0.95, indicating high burstiness. The self-similarity of Ethernet traffic has important implications for traffic engineering, including the modeling of individual Ethernet users, the inadequacy of conventional notions of burstiness, and the effects on congestion management in packet networks. Self-similar traffic models are more accurate and realistic than conventional models, and further research is needed to fully understand their implications.Ethernet traffic is statistically self-similar, meaning it exhibits similar patterns across all time scales. This behavior is not captured by conventional traffic models and has significant implications for network design, control, and analysis. Self-similar traffic lacks a natural "burst" length, with bursts consisting of bursty subperiods separated by less bursty subperiods at every time scale. Analysis of hundreds of millions of Ethernet traffic measurements from 1989 to 1992 shows that self-similarity is evident, with the Hurst parameter indicating the degree of burstiness. Self-similar traffic is different from conventional telephone traffic and standard packet traffic models, such as Poisson or Markov-modulated Poisson processes. The burstiness of LAN traffic typically increases with more active sources, contrary to common assumptions. Self-similar models provide accurate and realistic descriptions of traffic scenarios in high-bandwidth networks. The paper discusses the implications of self-similar traffic for congestion control, showing that congestion management in self-similar environments differs from conventional models. The paper also presents statistical methods for analyzing self-similar data, including R/S analysis, variance-time plots, and periodogram-based methods. The results show that Ethernet traffic is self-similar with Hurst parameters ranging from 0.75 to 0.95, indicating high burstiness. The self-similarity of Ethernet traffic has important implications for traffic engineering, including the modeling of individual Ethernet users, the inadequacy of conventional notions of burstiness, and the effects on congestion management in packet networks. Self-similar traffic models are more accurate and realistic than conventional models, and further research is needed to fully understand their implications.