Rate/Size Dependency Project

TO DO: I am still working on the synthetic data sets. They should be ready for analysis in the next few days.

Guide: Scatter plots and EDA results are only for non-0 duration flows.
            Thresholding lines for L-L Correlation plots: 1KB, 10KB, and 100KB for Size.
                                                                       0.01sec, 0.1sec, 1sec, 5sec, 100sec for Duration.
            Thresholding numbers for EDA: Top 100, 200, 500, 1000, 2000 (in the table), 5000, 10000, and 20000.
            Thresholding percentages for EDA: Top 0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 1, 2, 5, 10, 20 (%).

            Bell Lab "all" and "rest" data sets are split into two subsets to avoid calculation problems of Matlab.

Note: Some plots do not show thresholding lines properly especially in cases of large data sets.
           However, the printed plots look all right.

Connection

Data Units

Gaussian Simulation

Summary

EDA Summary plots: Size-Rate Size-Duration   Duration-IRate
                                 (blue: All, magenta: HTTP, red: SMTP/NNTP, green: Rest, solid: CONN, dotted: DATA_UNIT)
EDA Summary plots (based on percentage): Size-Rate  Size-Duration    Duration-IRate    Size-IRate
                                                                   Synthetic data

1. Log-Log correlation plots of Size-Rate show the same story (contradiction in two regions).
    Roughly, CONN: 0.8 vs. 0.4, DATA_UNIT: 0.9 vs. 0.1.
2. Synthetic all (DATA_UNIT) looks different (strong linear relationship in Size-Duration and Duration-IRate).
3. Connection level is more dependent: DATA_UNIT < CONN.
4. Abilene data are most dependent in Connection level: UNC, Bell < Abilene.
5. Abilene data are most independent in Data unit level: Abilene < Bell < UNC.
6. SMTP/NNTP is most dependent and All and HTTP are most independent: All, HTTP < Rest < SMTP/NNTP.
7. In Size-IRate, all of them are quite independent.