Daniel Zappala

Network Neutrality

Tue 18 November 2014 network neutrality

We are covering network neutrality in CS 660 this week. This is a great topic for exploring the structure of the Internet and business relationships among ISPS, as well as a chance to grapple with difficult policy questions.

Network neutrality has been in the news a lot lately, with the FCC considering whether to impose regulations. A number of articles from Vox cover this space for the general public:

Network Neutrality Inference

The first paper we read is from SIGCOMM 2014, titled Network neutrality inference [1]. This paper constructs a model to show the conditions where it is possible to detect network neutrality violations, using only measurements of application performance from endpoints. The model is based on a clever insight from tomography, where inferences about network performance are made using a linear system of equations. The authors make the observation that network neutrality violations result in an unsolvable system of equations. They use this idea to show that they are able to isolate a single link or a sequence of links as causing a network neutrality violation.

The following figure from the paper illustrates the model:

neutrality inference example network

In this network, three applications are using paths p1, p2, p3. The ISP separates the traffic into two separate classes, and treats traffic from p2 as having lower priority than traffic from either p1 or p3.

The authors show how to split the non-neutral link l2 into two pieces, where l + 2(1) captures the performance of the link for class 1 (p1, p3) and l + 2(2) captures the performance of the link for class 2 (p2). This is shown in the figure below:

neutral equivalent

Notice p2 is the only path that traverses link l + 2(2), so that any performance differences can be observed as a network neutrality violation.

The paper discusses the formal conditions for observability of network neutrality violations and identifiability conditions for links involved in a violation. Based on these results, it develops an algorithm that can have no false positives but no false negatives. The performance of the algorithm is verified using emulation on a variety of topologies.

ISP Interconnection and its Impact on Consumer Internet Performance

We next paper we read is a technical report from the Measurement Lab Consortium, called ISP Interconnection and its Impact on Consumer Internet Performance. This report examines measurements between access ISPs (AT&T, Comcast, Centurylink, Time Warner Cable, and Verizon) and transit ISPs (Cogent, Level 3, and XO) in a variety of locations around the United States. They observe performance degradations such as higher round trip times, increased packet loss, and decreased throughput during periods of high network use among certain pairs of access and transit ISP. The evidence indicates that this is due to business relationships between ISPs.

A good example of the problems observed is shown below:

download throughput for Cogent with three ISPs

This shows the throughput between New York City customers of TWC, Comcast, and Verizon as they connect through the Cogent transit ISP. Notice that degradation is severe betwen June 2013 and February 2014. It is well below 4 Mbps and as low as 1 Mbps. During this same period, throughput between New York City customers of Cablevision and Cogent was around 14 Mbps. Likewise, performance between customers of all four ISPs and the Internet transit ISP had an average throughput of 14 Mbps during the same time:

download throughput for Cogent with Cablevision

This indicates the problem is the connection between specific access ISPs and transit ISPs. The consequence of this problem is nicely illustrated below, where daily download speeds fall below 0.5 Mbps.

daily download throughput

The report has numerous additional details on other metrics such as round trip times and loss rates, as well as data from other areas of the country. All the data and tools are open source.

References

[1]Zhiyong Zhang, Ovidiu Mara, and Katerina Argyraki. 2014. Network neutrality inference. In Proceedings of the 2014 ACM conference on SIGCOMM (SIGCOMM '14). ACM, New York, NY, USA, 63-74. DOI=10.1145/2619239.2626308 http://doi.acm.org/10.1145/2619239.2626308
[2]Measurement Lab Consortium. 2014. Technical Report. http://www.measurementlab.net/static/observatory/M-Lab_Interconnection_Study_US.pdf
UtahJS Talk on Civic Hacking
Civic Engagement Workshop