estimating packet loss rate in the access through application
TRANSCRIPT
ACM SIGCOMM Workshop on MeasurementsUp and Down the Stack (W-MUST) 2012
Estimating Packet Loss Rate in the Access Through Application-Level
Measurements
*Nexa Center for Internet & Society, DAUIN, Politecnico di Torino+DET, Politecnico di TorinoǂDAUIN, Politecnico di Torino
S.Basso* M.Meo+ A.Servettiǂ J.C.De Martin*
Helsinki, 17 August 2012
17 August 2012 Simone Basso – Politecnico di Torino 2
Rewind: Nexa Center and earlier studies on network neutrality
● Nexa Center: academic research center that studies the Internet from a multidisciplinary point of view (technology, law, economics)
● January 2009 – legal analysis of 10 wired and 4 wireless Italian broadband ISPs contracts (http://bit.ly/GQ6FEK)
● Decision to develop a software research project to monitor network neutrality: Neubot (http://neubot.org)
17 August 2012 Simone Basso – Politecnico di Torino 3
Neubot: the network neutrality bot (active application-level tests)● Neubot is a daemon that runs in the background and
periodically performs active tests towards servers provided by Measurement Lab (http://measurementlab.net)
● Tests probe the end-to-end path between the user and the server (including at least access and home network), aim at saturating the bottleneck, run for a short amount of time (5-10 s), emulate HTTP and BitTorrent, and perform goodput and RTT measurements at application level
● Results are centrally collected, providing an application-level characterization of what one can see, from the user vantage point, using different protocols
17 August 2012 Simone Basso – Politecnico di Torino 4
Rationale and issues of application-level measurements
● Application-level tests are appealing because they are simple, scalable and portable to other contexts
● Unfortunately, a lot of very useful packet-level information is missing (e.g. retransmissions and other TCP events)
● Of course, one cannot just compare (as in “which one is better from the point of view of a human?”) two measurements by the goodput alone, because the goodput is inversely proportional to the RTT
17 August 2012 Simone Basso – Politecnico di Torino 5
Idea: estimate packet loss rate and use it for comparison
● We propose to invert the Mathis formula, to estimate the packet loss rate (PLR) from RTT, goodput and maximum segment size (MSS)
● In turn, the comparison between two measurements will be performed in the bidimensional (RTT, PLR) space
● The question is whether the estimation is sufficiently robust, given that we are not working in ideal conditions (e.g. short tests, simple RTT estimation)
17 August 2012 Simone Basso – Politecnico di Torino 6
Step #1: evaluation in atestbed environment
17 August 2012 Simone Basso – Politecnico di Torino 7
Testbed experiments results
17 August 2012 Simone Basso – Politecnico di Torino 8
Step #2: evaluation with controlled Internet experiments
17 August 2012 Simone Basso – Politecnico di Torino 9
Controlled experiments results
17 August 2012 Simone Basso – Politecnico di Torino 10
Step #3: evaluation in testbed emulating an ADSL connection
17 August 2012 Simone Basso – Politecnico di Torino 11
Results of testbed emulatingan ADSL connection
17 August 2012 Simone Basso – Politecnico di Torino 12
Conclusion and future work
● It seems good enough for coarse grained classification of application level performance
● The estimated PLR depends on the connection speed (so that a “better” connection for the user gets a lower PLR estimate)
● The estimated PLR still depends on the RTT: does running tests with higher RTT for longer times mitigates the problem?
● Does the model work reasonably for request-response application-level patterns?
17 August 2012 Simone Basso – Politecnico di Torino 13
Thank you for your attention!
Simone Basso ([email protected])Nexa Center for Internet & Society (http://nexa.polito.it/)
Dept. of Computer and Control Engineering (DAUIN)
Politecnico di Torino, Italy
The Neubot project (http://neubot.org/)
http://twitter.com/neubot
http://facebook.com/neubot
17 August 2012 Simone Basso – Politecnico di Torino 14
Appendix: Mathis formula
● MSS is the maximum segment size● C is 0.93 for random losses and delayed ACKs● RTT is the round trip time● PLR is the packet loss rate