Meredith Whittaker is Open Source Research Lead at Google.
For much of 2013 and 2014, accessing major content and services was nearly impossible for millions of US Internet users. That sounds like a big deal, right? It is. But it's also hard to document. Users complained, the press reported disputes between Netflix and Comcast, but the scope and extent of the problem wasn't understood until late 2014.
This is thanks in large part to M-Lab , a broad collaboration of academic and industry researchers committed to openly and empirically measuring global Internet performance. Using a massive archive of open data, M-Lab researchers uncovered interconnection problems between Internet service providers (ISPs) that resulted in nationwide performance slowdowns. Their published report, ISP Interconnection and its Impact on Consumer Internet Performance , lays out the data.
To back up a moment—interconnection sounds complicated. It's not. Interconnection is the means by which different networks connect to each other. This connection allows you to access online content and services hosted anywhere, not just content and services hosted by a single access provider (think AOL in the 1990’s vs today’s Internet). By definition, the Inter-net wouldn't exist without interconnection.
Interconnection points are the places where Internet traffic crosses from one network to another. Uncongested interconnection points are critical to a healthy, open Internet. Put another way, it doesn't matter how wide the road is on either side—if the bridge is too narrow, traffic will be slow.
M-Lab data and research exposed just such slowdowns. Let’s take a look…
The chart above shows download throughput data, collected by M-Lab in NYC between Feb 2013 and Sept 2014. The reflects traffic between customers of Time Warner Cable, Verizon, and Comcast—major ISPs—and an M-Lab server hosted on Cogent's network. Cogent is a major transit ISP and many content and services are hosted on Cogent’s network and on similar transit networks. Traffic between people and the content they want to access has to move through an interconnection point between their ISP (TWC, Comcast, and Verizon, in this case) and Cogent. What we see here, then, is severe degradation of download throughput between these ISPs and Cogent that lasted for about a year. During this time, customers of these three ISPs attempting to access anything hosted on Cogent in NYC were subjected to severely slowed Internet performance.
But maybe things are just slow, no?
Here you see download throughput in NYC during the same time period, for the same three ISPs (plus Cablevision). The difference: here they are accessing an M-Lab server hosted on Internap's network (another transit ISP). In this case, in the same region, for the same general population of users, during the same time, download throughput was stable. Content and services accessed on Internap's network performed just fine.
Couldn't this just be Cogent's problem? Another good question…
Here we return to Cogent. This graph spans the same time period, in NYC, looking again at download throughput across a Cogent interconnection point. The difference? We’re looking at traffic to customers of the ISP Cablevision.
Comparing these three graphs, we see M-Lab data exposing problems that aren't specific to one ISP or ISPs, but a problem with the relationship between pairs of ISPs—in this example, Cogent when paired with Time Warner, Comcast, or Verizon. This relationship manifests, technically, as interconnection.
These graphs focus on NYC but M-Lab saw similar patterns across the US as researchers examined performance trends across pairs of ISPs nationwide—e.g., whenever Comcast interconnected with Cogent. The research shows that the scope and scale of interconnection-related performance issues were nationwide and continued for over a year. IT also shows that these issues were not strictly technical in nature. In many cases, the same patterns of performance degradation existed across the US wherever a given pair of ISPs interconnected. This rules out a regional technical problem and instead points to business disputes as the cause of congestion.
M-Lab research shows that when interconnection goes bad, it’s not theoretical: it interferes with real people trying to do critical things. Good data and careful research helped to quantify the real, human impact of what had been relegated to technical discussion lists and sidebars in long policy documents. More focus on open data projects like M-Lab could help quantify the human impact across myriad issues, moving us from a hypothetical to a real and actionable understanding of how to draft better policies.
