For researchers, policy makers, governing bodies, advocacy groups, or anyone who wants to understand M-Lab data and how it compares to other internet measurement data sets, M-Lab has begun to develop recommendations for appropriate use of our data in their analyses and reports.
Measurement Lab’s mission is to measure the Internet, save the data, and make it universally accessible and useful. Our goal is to measure a user’s whole experience of the Internet and how effectively it supports them day to day. To improve the Internet, we need to model how it behaves for the user in their everyday lives. Furthermore, these models should be built on transparent and verifiable data and methods to ensure that users, policy makers, advocates, researchers and anyone else with an interest in the future of the Internet, can make decisions with the same information.
- Different goals are best served by different data sources. No single source is best for all purposes. Many insights arise only when comparing and contrasting multiple data sources.
- The Internet is the interconnection of many independent networks. The portion of this topology that is measured is a function of client and measurement server placement.
- People measuring to a server within their access ISP’s network will assess only that ISP’s network, not the user’s connection to the Internet. This server placement is called “on-net”
- People measuring to a server within a transit provider that “peers” with the access ISP will assess the approximate path to content outside of the access network. Because the path includes an interconnection, this measurement may also assess the quality of the peering between the access ISP and transit provider. This server placement is called “off-net”
- The FCC MBA Program Technical Index notes, “The use of both on-net and off-net nodes provides a measure of confidence in the test results.”
- M-Lab data is an excellent source for researchers who want:
- open data and transparent methodology
- to measure the interconnections between independent networks (the “inter” part of the Internet)
- full access to extensive, detailed data, rather than broad generalizations i.e.:
- break down by ISP, time of day, day of week, IP subnets
- break down across different destination transit providers.
- Compare M-Lab data to other data sources when researchers:
- want to measure the last mile connection
- are interested in exhaustive data on a specific ISP and
- have access to (and trust) proprietary internal data
- using proprietary data works for your needs
- need controlled or verified client demographic, connectivity, or geolocation data
- When analyzing and illustrating data:
- Don’t oversimplify
- Aggregate by ISP in addition to time/date and location
- Be aware and illustrate multimodal distributions
- Use histogram and logarithmic scales
- Take into account, and compensate for, client bias and population drift
- Don’t oversimplify
As with any data-driven research initiative, researchers using M-Lab’s NDT dataset should first consider whether or not our methodology meets the expectations of their use case. Measurement services on the M-Lab platform are an excellent source if you are looking to measure the “inter” part of the Internet or the interconnections between providers. By running an exclusively off-net platform, M-Lab measures more than the last-mile network by design. M-Lab data will work best for researchers who are looking to represent the full path from a user to the backbone of the Internet. Complementary datasets such as Ookla’s speedtest.net provide data from servers as near in network distance to the user as possible, likely within the user’s access network, and are better suited for researchers looking to only measure the last-mile connection.
M-Lab data is freely available, open source, large, and longitudinal. Anyone can sign up to gain access to the data via BigQuery at no cost. All of the software that runs our platform, pipeline and measurement services (i.e. all of the code that generates M-Lab data) is open for review on GitHub. Users have access to the full history of data, as well as pre-filtered views that only show tests that meet our team’s current, best understanding of test completeness and research quality. By providing open source for all components, M-Lab offers an exhaustive view into the provenance of all collected and published data.
Also, researchers are encouraged to use M-Lab’s data as a public benchmark to compare other performance datasets. For example, if a researcher has proprietary performance data for a single ISP or content provider, they can compare this to open NDT data to assess whether their clients perform at least as well as M-Lab clients from similar network or geographic locations. Or, a network provider could compare their known topology to open traceroute data to compare predicted and actual routes.
Keep in mind that mean and medians summarize more complex underlying distributions. If the underlying distribution is a single narrow peak, interpretation is easy, but if it is multimodal, or has long tails or high skew, it is less obvious how to interpret a mean or median.
Multimodal distributions usually appear because there are several distinct populations in the underlying data. Breaking down data on dimensions like ASNumber, peak/off-peak, or urban/rural will often allow for clearer interpretation.
In particular, individual ISPs may serve different market niches, implement services using different technologies, or have differing levels of infrastructure investment. Different ISPs operating in the same geographic area may therefore have very different performance characteristics.
Peak/off-peak distinctions are often useful to identify regions or ISPs where peak traffic exceeds some internal infrastructure capacity. Peak hour end-to-end throughput may not match last mile throughput, or off-hour throughput, due to queuing at internal switches during peak traffic times.
Use histograms and logarithmic scales
Histograms can clearly illustrate more complex distributions, with long tails, or multiple groupings of fast or slow clients. When providing geographic visualizations, a complementary histogram can provide greater detail about a single geographic area.
Internet performance ranges over many orders of magnitude, and, just as sound levels are measured in dB, internet performance is best represented on log scales. Using linear scales or linear averages over-weights the fastest measurements (or the largest MinRTT values).
If you do wish to calculate averages, and have done your due diligence to account for factors like client bias, sample size, etc. then the average should be calculated in logarithmic space (also known as the geometric mean), and histograms should use logarithmically spaced buckets, usually with 2 to 5 buckets per order of magnitude depending on how many samples are available and how much detail is useful.
Take into account, and compensate for, client bias and population drift
Everything changes - sometimes all at once. Trends in the data evolve from a combination of:
- Internet evolution
- Client population changes
- M-Lab platform or service changes (e.g. shifting from Cubic to BBR congestion control)
In aggregate, any change (to the Internet, client population, or service software) may lead to improving or declining performance measurements. For example, improvement in an area may mean the network is getting better, or it may mean a few very fast clients are running a lot of tests, or that households with fewer options are being squeezed out. Declining overall performance measurements may mean an oversubscribed interconnection point, or actually reflect growth in service in low income markets or rural areas where service is slower.
Researchers should look at client population changes when comparing statistics across different time periods (whether means, medians, or histograms) 1. Aggregate statistics may reflect actual network evolution, or may be dominated by growth in a specific population of clients (e.g. from one fast growing ISP), or changes in the testing behavior of existing clients (e.g. through increased awareness or automated testing).
[ 1 ] The most common way client populations have been identified in research has been using the IP address. This is an imperfect method, and both our team and other researchers are actively working to develop other methods that can be more precise.