...
Ookla maintains a fleet of speed-testing servers all over the world. A user chosen server chosen server (by default selected as the nearest server to the client, though the client/user can select their own server) downloads a file of incompressible data to the user's client and then uploads it. They calculate the throughput by knowing the amount of data transferred and then dividing this by the time taken to transfer the data in each direction. A summary of this data is provided to the client user in real-time showing average download speed, upload speed, and ping Round Trip Time (RTT).
...
The results from Ookla are mainly oriented to broadband access to for the residence and the performance of the local ISP. It broadband providing ISP. The measurements are made at irregular intervals as determined by the whims of the client. The clients are self selected. Ookla also measures the true throughput separately in both the up- and downstream directions. PingER on the other hand is more oriented to measurements beween academic and research sites, the measurements are made at regular intervals (every 30 minutes) and the clients are selected by the server administrators. PingER also derives the throughput from its RTT and loss measurements using the Mathis formula (TCP throughput ~ 1460*8/(RTT*sqrt(loss)) where the result is in kbits/s if the RTT is in milliseconds. Thus the measurement is an average of the up and downstream directions. Though the Mathis formula is for classic versions of TCP's (e.g. New Reno) congestion control as used in most off the shelf computers today, it assumes TCP loss (and, and TCP provokes loss to detect congestion) whereas PingER gets the loss from pings and thus only samples loss. For these reasons and others PingER is only a rough estimate of throughput.
Our Ookla data has about 19 million IP client addresses. It is unclear how many times a given client makes a measurement or if it uses different servers from the default. PingER has about 1900 monitoring host (server) remote host (client) pairs each measured regularly every 30 minutes, or about 3 million measurements (where a single measurement is of 10 pings) per month.
To validate whether the order of magnitude of the throughputs measured by Ookla agreed with those derived from the PingER losses and Round Trip Time (RTT) measurements we plotted the Ookla speedtest download and average measurements versus the Normalized Derived throughputs (NDT) from Pinger. It is seen that the order of magnitudes agree and there is a good positive correlation (R2 ~ 0.53). To see how well the Ookla speedtest download results agreed with those of the ZDnet's Australia Speedtest (though the name Speedtest appears in both they are not related) we also plotted the two speedtests against one another. As a comparison we also show the ZDnet Speedtest results vs. the PingER NDT.
...