Throttling is a notoriously difficult occurrence to measure. Why?
By Lai Yi Ohlsen, Director, Measurement Lab, a fiscally sponsored project of Code for Science & Society
Throttling is difficult to define in a data-driven way
Access Now defines bandwidth throttling as: “the intentional slowing of an internet service or a type of internet traffic by an internet service provider (ISP). It is employed by ISPs to regulate network traffic and ease bandwidth congestion. When one’s internet bandwidth is throttled, it results in poor performance of web content or service for the user.” According to this definition, unlike the binary nature of shutdowns, the behavioral patterns of throttling are best observed on the spectrum of Internet performance. “Slow” is a relative term, and “intentional” behavior can often hide under the guise of unintended consequences. These two implicit concepts make throttling difficult to apply a unilateral, metric-based definition that can apply across all types of regions and contexts.
For one thing, the behavior of ISPs is difficult to confidently explain from the outside, even without the potential involvement of nefarious actors. Sometimes poor performance is unexpected and is simply a result of inadequate network management. But other times it is an expected behavior that is intended to level out traffic across networks to prevent disproportionate performance. Both behaviors are ostensibly benign, but produce comparable measurement results, making it complex to define the difference.
In the Internet Freedom and censorship detection context however, throttling often implies not only the involvement of an ISP but also the participation of a government or state actor whose intention is to decrease citizens ability to participate online. Potential political intention only complicates the interpretation of Internet performance measurements, as it just adds another vector of uncertainty when analyzing the data. The ability for throttling to be attributed to a number of different causes gives state actors an optimal tool to obfuscate their behavior and avoid accountability.
Context is key
Though throttling is not always immediately distinguishable, it is possible to learn from the context of the region, socially, politically and technically. For example, if a region’s connectivity consistently shows poor performance during peak traffic time, this behavior will be less indicative of throttling if it occurs during a political event, such as an election. However, if the region typically demonstrates strong performance, and it noticeably weakens during a political event, it could be an indication of intentional interference. Similarly, if every Autonomous System (AS) shows evidence of poor performance, except for the AS run and owned by the state government, it might suggest potential political preference. Studying routing data in relation to performance data could also reveal preference for content that is hosted inside of a country vs. outside. All of these examples and more require an intimate understanding of how a region’s Internet behaves “normally”, as throttling inherently implies a departure from these conditions. Such an understanding requires an interdisciplinary approach, as the behavior of the Internet can often be related to how the social and political climate has changed over time. Internet measurement data can reveal more when combined with a local lens.
So what do we do?
The good news is, we have data to work with. Between RIPE Atlas, OONI, M-Lab, CAIDA, Google’s Transparency Report and others, there is a solid amount of open Internet Performance data. The open question is how do we use it? There is of course no easy, immediate answer but any successful solution will be interdisciplinary, coordinated, and context-specific. Throttling exploits an inherent limitation of Internet measurement: data can tell us what is happening but it does not always tell us why. But if researchers can think of each dataset as a point in the direction and combine those hints with what experts know about the relationship between Internet and politics in regions where throttling has reportedly occurred, we can begin to identify trends and patterns. Measurement of throttling has the potential to be a key component in the movement to hold governments accountable for censorship of their citizens. At the end of the day, the symptoms of Internet interference are often felt by people first. We just need to identify the data that will validate their experience.