Optimizing performance in heavy-tailed system: a case study

Heavy-tailed distributions play an important role in network analysis. Resent researches have highlited that this kind of distributions can describe a behavior of some real network process better than others (e.g. exponential distribution). Theoretical analysis and experimental measurements have shown the presence of heavy tails in a system process lead to burstiness: the process has bursts, which is much bigger than ordinary value of the process. This effect has a negative influence on system characteristics: performance, reliability, quality of service. We discuss main statistical properties of heavy-tailed distributions and some practice examples, which illustrate evidence of heavy tails in network systems. The main goal of the work is to find and describe ways, which can minimize the effect of heavy tails on network performance.