When thinking about broadband in the United States, the first thing people likely think about is whether their connection is fast enough– are they getting the speeds they need to do business, go to school, etc. The dominance of “speed” in assessing broadband service goes all the way to the top- the FCC defines broadband according to the achievable download and upload speeds to the Internet. But generic speed test measurements only go so far in observing a connection’s performance, and M-Lab and the research community are working to expand the concept of broadband measurement beyond basic speeds.
To illustrate why we’re suggesting new approaches to measurement are needed, let’s take a look at some recent examples of national level aggregate data that communities are using to inform their planning for new infrastructure dollars. One starting point is the NTIA’s Indicators of Broadband Need map, which we wrote about this summer. Our goal with that post was to provide context and provenance of the different data sources presented in the map, since each represents a different type of measurement of an Internet connection. While we are pleased to have NDT data from M-Lab represented in the map, there are some issues with the presentation that could make this data confusing at a basic level, and at worst could lead to incorrect conclusions.
Our post from July 2021 discussed the various data sources in the NTIA Indicators map in relation to the FCC’s current definition for a broadband connection– one that advertises 25 Mbps download and 3 Mbps upload speeds to the Internet. In most ISP’s terms of service, those speeds are the maximum possible speeds, not those that we should expect to receive 24/7/365. So when we look at the NTIA map, we can easily compare some of the most common broadband data sources, but there are some questions that are worth exploring further:
- Why are median upload and download speeds presented instead of the average maximum speeds?
- What are the speeds broken down by provider in the map? Why aren’t those included?
- The NTIA map provides aggregate data for one six month period, and they plan to provide additional 2 quarter time periods following the FCC’s Form 477 release cycle. Considering that networks likely change more frequently, why are other, smaller time periods like quarter or month not available?
NTIA’s Data Source for NDT Statistics
NTIA is using aggregate NDT data from M-Lab using an approach very similar to our own Statistics Pipeline Service, which provides aggregate statistics for NDT tests which follow our team’s recommendations for working with NDT data. The statistics can be accessed using an API, or within our public BigQuery tables in the dataset,
measurement-lab.statistics. For each day, we provide a histogram of measured download and upload speeds, and overall daily statistics. As of this writing, the statistics pipeline service provides this data from Jan. 1, 2020 to present.
Including other metrics than median
What people are really wanting to compare when they look at the NTIA map is whether their locale is meeting the FCC’s broadband standard in light of the agency’s Form 477 reported data about availability of service. But when we’re talking about the standard and “getting what you paid for”, the map may lead to wrong assumptions.
The advertised speeds of providers’ service tiers are not guaranteed, but are the speeds up to which your service can perform. As we explained in our post, an Ookla speed test most closely aligns with the advertised speeds that your connection may reach. NDT is a diagnostic test that measures using a single TCP stream. The single stream “speed” metric doesn’t measure the connection’s link capacity, but rather how well a single TCP stream can take advantage of that capacity. While M-Lab researchers have suggested that the ndt7 protocol may be able to also provide a link capacity metric, Ookla is currently more suitable to directly measure link capacity.
If I were to run an Ookla test a few times a day over a month or two from my home network, I could have a fairly accurate representation of upload and download speed that is tied to the service that my ISP delivers in their last mile network 1. I could also run an NDT test to see an accurate representation of TCP’s performance to the geographically nearest M-Lab server outside of my ISP’s network. I could look at the trend in tests over time and might look at the average of maximum measurements for each day over the two months.
But the example above is only useful to me as an individual- I know my ISP, connection plan, cost, etc. Which is why reporting something other than median speeds in the NTIA tools would be helpful. For example, the query NTIA is using includes the minimum, median, and maximum aggregate upload and download measurements, but also includes the 10th, 25th, 75th, and 90th percentiles. M-Lab’s stats-pipeline currently provides the 25th and 75th percentile aggregations, and we’ll be adding others in upcoming releases to support data integrations like the NTIA directly in the future.
While M-Lab has also published median speeds from NDT, we no longer recommend looking at only median or average in the context of advertised broadband speeds. In the past, M-Lab used the median metric because of its use within the FCC’s Measuring Broadband America Program (MBA) reporting of (among other metrics) the ratio of weighted median speeds to advertised speeds for participating ISPs 2. The median of measurements can be representative of the performance achieved by a sampling of tests, but while researchers can infer the speed tiers of testers over a sample of NDT tests, M-Lab doesn’t know directly the tiers of service for people running them. Additionally, crowdsourced measurements (NDT and those of other platforms), while having large volumes of tests are also more challenging to use when selecting a representative sample. Unless we are collecting NDT data from known clients, as is possible with Murakami, a representative sample can be difficult to gather.
Aggregation by provider
In its current form, the NTIA is providing aggregate speeds from NDT and Ookla for geographic areas. While this might give us a general idea of Internet service as measured by these tools in a county, state, or territory, it also is not very specific. Including aggregation by provider in each geography and time period would better represent the service of individual ISPs instead of all of them together.
While aggregating by provider is desirable, it is not available in NTIA’s public map. We don’t have access to NTIA’s private tool for state and tribal broadband offices, but Ookla data at the provider level may be available within it since Ookla does market this data in their product offerings.
M-Lab publishes aggregate data by provider by default, using the Autonomous System Number and Name (ASN) of the company using each IP address at the time tests are run. While NTIA doesn’t include this aggregate by ASN in their tools, it is available in our statistics pipeline, and presented on our available DataStudio reports.
M-Lab recommends that analyses always include breakdown by ASN for NDT data, unless there is a specific reason to not do so, since it more accurately reflects the measured performance of each individual provider. It’s worth mentioning that we have started to decline support requests that contradict these recommendations.
In 2020, an organization asked M-Lab to provide aggregate data for regions within countries in the European Union for the year 2019. The request was for aggregate NDT data for each region over the entire year without breakdown by ASN or provider. Because we did not think this would accurately represent either the state of broadband service as measured by NDT, we declined to continue contributing data to the initiative. We have also declined requests to endorse reports that rank countries across the world by the speeds that NDT measures. Our data remains publicly available, so these organizations could perform their own queries and analysis if they desire, but we strongly encourage the use of the appropriate methods to ensure analyses and conclusions are sound.
Aggregate by smaller time periods
The NTIA tools provide NDT and other data sources aggregated in a single 6 month time block: January 1, 2019 through June 30, 2019. NTIA also plans to provide additional snapshots of aggregate data over additional time periods, following the release pattern of the FCC’s Form 477 data. While it makes sense to look at the same time periods for all data sources, being able to select a different date range for data sources that aren’t bound to the twice annual Form 477 release can provide deeper insights. Networks can change a lot over 6 months, particularly if they’re being upgraded or when infrastructure investments are getting rolled out. Being able to slice data collected before and after such events would be helpful to communities. Ookla’s public data is released in one quarter time periods, and more granularity is available from their product offerings. M-Lab makes all individual NDT test data available, so literally any time period is possible.
While M-Lab encourages research with our data and with other broadband data sources, we also are increasing our capacities internally to support and conduct research; providing clearer analyses and recommendations for people conducting analyses with our data; explaining our test results and providing context to end users who conduct tests; and engaging with the research community to determine gaps in measurements that would better inform an ecosystem of Internet measurement research.
We know people will continue to use M-Lab data for their analyses, and want to support them as best we can. As a start to that, we’ll be following up on many of the ideas presented here, and on our initial recommendations for analysis published last year, with a series of posts on how to approach working with our crowdsourced NDT data. Our research team is also working on new methods to improve the NDT statistics provided in our statistics pipeline.
An ndt7 test will use BBR. You can read more about BBR, congestion control algorithms and the evolution of NDT here: https://www.measurementlab.net/blog/evolution-of-ndt/#evolution-of-ndt ↩