The paper quoted that, in mid-1990s to early-2000s, most traffic is web while the amount of P2P traffic became comparable to web since then. This paper is to measure the P2P traffic, as well as web traffic for the knowledge of traffic models. The measurement is made on an Internet link of Waseda University (早稻田大學) for a night (6pm to 12mn).

The result is that, around 70% of flows are WWW and P2P flows accounts for less than 1%. But the traffic volue of WWW and P2P are respectively 58% and 22%. The flows’ interarrival time are exponential in both WWW and P2P traffic, even their mean interarrival time are respectively 0.015s and 15s. Their correlation coefficient against exponential is greater than 0.99.

Power spectral density is also obtained. The PSD is defined as

\[I(f)=\dfrac{1}{2\pi N}\sum_{k=1}^N T_\delta(k)e^{jkf/2},\]

where \(T_\delta(k)\) are the flow interarrival times. The PSD plot shows that, in WWW traffic, \(I(f)\) has a decrease tendency with \(f\) but P2P traffic remains flat with \(f\). This means WWW traffic has long range dependency but P2P does not have time correlation.

Considering the flow size distributions (number of bytes), the cummulative percentage-flow size plot is linear in the log-log scale when we consider the size of 10K to 100M. This means the flow size distribution can be modelled by Pareto distribution. The small flow size region increases slowly and the large flow size region increases rapidly in the cummulative percentage The Pareto parameter \(\alpha\) is measured to be 1.11 in the WWW traffic and 0.38 in the P2P traffic. Since \(\alpha<1\) in P2P, it is highly heavy-tailed. But the WWW traffic has finite mean in flow size.

The flow duration distribution of both WWW and P2P shows a non-linear behaviour in the log-log plot of cummulative fraction against duration. The nonlinearity suggests log-normal distribution, i.e. duration \(\log D \sim N(\mu,\sigma^2)\), which is found to be a very good approximate in the middle ranges. The log-normal distribution is also found in the flow rate (flow size divided by duration).

Bibliographic data

@article{
   title = "Flow Analysis of Internet Traffic: World Wide Web versus Peer-to-Peer",
   author = "Tatsuya Mori and Masato Uchida and Shigeki Goto",
   journal = "Systems and Computers in Japan",
   volume = "36",
   number = "11",
   pages = "70--81",
   year = "2005",
   note = "Translated from Denshi Joho Tsushin Gakkai Ronbunshi 電子情報通信學會論文誌, Vol. J87-D-I, No. 5, May 2004, pp. 561–571",
}