The second-order character of self-similar network traffic, i.e. its correlation existing at multiple time scales, has an enormous impact on network performance, which has been widely studied. In studying the impact of self-similar traffic on network performance and utilizing its correlation structure to design network control scheme and on-line inspection, a real-time estimate of the autocorrelation of network traffic is often necessary. For an effective and quick estimate, it is very useful to determine the required sampled data length according to the requirement of precision. In this paper, the relationship between Hurst parameter ( H), the precision of estimated autocorrelation and required sampled data length is discussed on the basis of fractional Gaussian noise (FGN) model and a simple calculating formula is proposed. Furthermore, a sharp increase is discovered in the variances of estimated autocorrelation with the same data length, when H>0.75. This is a new phenomenon we call ‘jumping burstiness’, which has not been reported before. This phenomenon shows that Hurst parameter could reflect the self-similar character of network traffic, but it is not enough to capture all the features. Experiments confirm that our results are not only valid to FGN model, but also able to offer reference to estimate the autocorrelation of other self-similar models, such as fractional auto-regression integrated moving average. At the same time, trace driven simulation shows that the existence of jumping burstiness in traffic has a remarkable impact on queueing performance.