Grid'5000 experiment

Jump to: navigation, search

TCP traffic self-similarity under loss (Networking)

Conducted by

Paulo Goncalves, Matthieu Imbert, Patrick Loiseau, Pascale Primet


Over the last decade, many research efforts have been devoted to the study of aggregated traffic time series collected at the core of networks. The pioneering works by Paxson and Leland showed that the Poisson hypothesis, which is relevantly used in phone networks, was not suitable to describe computer networks. Instead, self-similarity was proved a much more appropriate paradigm. Then, the theoretical work from Taqqu and collaborators identified the heavy-tailed nature of the file size distribution as a possible origin for the observed self-similarity. In addition, it gave the exact relation between the self-similarity index and the tail index that should be observed when the sources behavior is modeled with the ON/OFF model. Despite a controversial debate on the question, it has then been more recently stated that the TCP congestion control mechanism cannot be responsible for the self-similarity observed in the large time scales. On the opposite side, we show in this work that when the file size is heavy-tailed, the TCP congestion control mechanism under sufficiently high loss can annihilate the self-similarity that would be observed without any loss. For this work, we use large scale controled experiments performed on Grid5000. Independant TCP sources send files in an ON/OFF scenario with a heavy-tailed ON periods; and a constant loss rate is created via UDP cross traffic.


in progress


  • Nodes involved: 500
  • Sites involved: 2
  • Minimum walltime: 4h

Tools used

No information


Not yet

Shared by: Paulo Goncalves, Matthieu Imbert, Patrick Loiseau, Pascale Primet
Last update: 2009-03-02 14:18:07
Experiment #557