Accuracy of Time-Domain Algorithms for Self-Similarity: An Empirical Study

Self-similarity plays an important role in the performance analysis of modern computer networks. An important problem is then to obtain an accurate inference of the degree of self-similarity and use this value for design and control purposes. Several algorithms for inferring the degree of self-similarity in a time series are currently in use. Unfortunately, several variables affect the accuracy of these algorithms. In this paper we identify these sources of inaccuracies and find the correct values for obtaining minimum biased estimates of the parameter of self-similarity. This "tuning" is done to several time-domain algorithms for self-similarity. The effect of the series length in the accuracy of these algorithms is also studied. This is done by the use of a cumulative analysis of self-similar traces. Based on this study we propose the minimum length series to obtain accurate estimates of the self-similarity parameter