Studies On The Internet Measurement And Modeling

The development of Internet, especially the occurrence and popularity of new Internet applications, requires us to carry out Internet measurement and modeling. In this thesis, we focus on four aspects related to Internet measurement and modeling, including residential Internet user behavior, identification of peer-to-peer traffic, the peer-to-peer sharing objects popularity and the behavior of the instant message application. Main contributions of this thesis are outlined as following.1. We perform a detailed study on the residential Internet user behavior based on a RADIUS authentication and accounting data set taken from a capital city. On the aspect of the user activity, our results indicate that the users' usage time and the users'traffic both show serious inequalities and that there is a certain linear relationship between them. Our results also indicate that there is a significant asymmetry between the users' upload and download traffic and a significant linear relationship between them. Furthermore, we propose that the users' usage time and the user' traffic could be accurately approximated by a Gamma distribution and a Weibull distribution respectively. On the aspect of the session, we find that over30percent of the sessions are terminated due to the loss of carrier. Our results indicate that the session arrival process is a non-homogeneous Poisson process. Furthermore, we propose that the session duration could be approximated by a Weibull distribution.2. We propose a peer-to-peer traffic identification method for high speed network using packet sampling and application signatures and implement it with Snort by developing a flow state differentiating preprocessor. Furthmore, we evaluate the efficiency, false negatives and false positives of the method. The method is supported to be efficient and accurate by identifying BitTorrent traffic. In addition, we present two schemes for high efficiency and low false positives.3. We propose an active measurement method for the BitTorrent sharing objects popularity and measure the sharing objects popularity of four BitTorrent trackers. Our results indicate that the objects sharing popularity and the objects download popularity both show serious inequalities and that there is a significant linear relationship between them. Our results also indicate that the objects sharing popularity could be both approximated by a Zipf distribution and a Zipf-Mandelbrot distribution with the latter being more accurate and that the objects download popularity could not be approximated by a Zipf distribution. Furthermore, we propose that the objects download popularity could be accurately approximated by a Zipf-Mandelbrot distribution.4. We perform a detailed study on the user activity, conversation and information propagation network of the MSN application based on a data set captured from the Beijing city backbone. On the aspect of the user activity, our results indicate that the MSN users show a daily characteristic, and most users use MSN only a few days in a week. On the aspect of the conversation, we propose that the characters in a chat message could be accurately approximated by a Log-logistic distribution and the characters in a conversation could be accurately approximated by a lognormal distribution. We also propose that the messages in a conversation and the conversation duration could be both accurately approximated by a Generalized Pareto distribution. On the aspect of the information propagation network, we find that there is a super swarm containing over50%users and that there are some users with large clustering coefficient. We propose that the degree could be accurately approximated by a lognormal distribution which indicates that the network is not a scale-free network. We also propose that the normal swarm could be accurately approximated by a Zipf distribution. Furthermore, we find that the super swarm is a small-world network.