Network Link Dimensioning based on Statistical Analysis and Modeling of Real Internet Traffic

Abstract

Link dimensioning is used by ISPs to properly provision the capacity of their network links. Operators have to make provisions for sudden traffic bursts and network failures to assure uninterrupted operations. In practice, traffic averages are used to roughly estimate required capacity. More accurate solutions often require traffic statistics easily obtained from packet captures, e.g. variance. Our investigations on real Internet traffic have emphasized that the traffic shows high variations at small aggregation times, which indicates that the traffic is self-similar and has a heavy-tailed characteristics. Self-similarity and heavy-tailedness are of great importance for network capacity planning purposes. Traffic modeling process should consider all Internet traffic characteristics. Thereby, the quality of service (QoS) of the network would not affected by any mismatching between the real traffic properties and the reference statistical model. This paper proposes a new class of traffic profiles that is better suited for metering bursty Internet traffic streams. We employ bandwidth provisioning to determine the lowest required bandwidth capacity level for a network link, such that for a given traffic load, a desired performance target is met. We validate our approach using packet captures from real IP-based networks. The proposed link dimensioning approach starts by measuring the statistical parameters of the available traces, and then the degree of fluctuations in the traffic has been measured. This is followed by choosing a proper model to fit the traffic such as lognormal and generalized extreme value distributions. Finally, the optimal capacity for the link can be estimated by deploying the bandwidth provisioning approach. It has been shown that the heavy tailed distributions give more precise values for the link capacity than the Gaussian model.