Strengthening measurements from the edges: application-level packet loss rate estimation

By: 
Simone Basso, Michela Meo, Juan Carlos De Martin
Appears in: 
CCR July 2013

Network users know much less than ISPs, Internet exchanges and content providers about what happens inside the network. Consequently users cannot either easily detect network neutrality violations or readily exercise their market power by knowledgeably switching ISPs. This paper contributes to the ongoing efforts to empower users by proposing two models to estimate -- via application-level measurements -- a key network indicator, i.e., the packet loss rate (PLR) experienced by FTP-like TCP downloads. Controlled, testbed, and large-scale experiments show that the Inverse Mathis model is simpler and more consistent across the whole PLR range, but less accurate than the more advanced Likely Rexmit model for landline connections and moderate PLR.

Public Review By: 
Nikolaos Laoutaris

Building upon their previous experience with Inverse Mathis (Inv-M) the authors propose Likely Rexmit (L-Rex), a new model for estimating the Packet Loss Ratio (PLR) of FTP like transfers. Unlike previous approaches that required access to kernel level TCP info, both Inv-M and L-Rex operate at the application layer by observing the timing and the return values of recv() calls. L-Rex was designed primarily for ADSL and fast ethernet networks. The authors present an extensive comparison between Inv-M and L-Rex in a controlled testbed as well as in the wild. For the latter, they integrate both models in the set of tests performed by the Neubot tool that is currently being used by approximately 1500 users around the world. The comparison reveals that L-Rex is more accurate in estimating the PLR for moderate and high values (loss rate above 1e-4). For low PLR, however, Inv-M performs better. The reviewers, all of which have developed related measurement tools in the past, were pleased by the final version of the paper and the effort that the authors put in simplifying the comparison between the two models and sending a clearer message to the reader regarding the circumstances under which the performance benefits of L-Rex become more pronounced. In their final round of reviews they recommended to the authors to describe more uses for L-Rex, other then obvious stand-alone. Specifically, how the model may be combined with additional measurements from other sources or be integrated with an actual application and provide a closed feedback loop regarding high level quality of experience.