[ckuehl]

For what it's worth, ISPs are definitely shaping traffic to try to game results on public speed tests.

I had an issue with Comcast last year where I would have large latency spikes for ~2 hours every night, and speedtest.net always remained at normal latency: https://i.fluffy.cc/kNCxLMbkF2wx7NHFJxbzcgX35Bsb5nvl.png

I couldn't find any non-speedtest sites where my latency was less than 100ms, but Comcast's own speed test, several other public speed tests, speedtest.net, and even ookla.com (the company behind speedtest.net) were perfect (the ~20ms you see on the graph).

[mcpherrinm]

Speedtest.net and other speedtests often choose the speedtest server that's close to you and has low ping.

What you describe is probably congestion in the network somewhere that isn't the last mile. Speed tests aren't a good indicator for any issues occurring anywhere but the last mile, as they won't take the same path with fairly high probability.

[gruez]

Given that he has a nice graph with hundreds of data points, I'm guessing he got the ping results from the ping command, not by running the speedtest.net test(which would actually select the lowest ping server)

[mcpherrinm]

You're right, though I had assumed he was pinging whatever server speedtest had selected, which may be erronous.

Really, my point was that congestion is often localized in a network. The fact that it isn't present in a path doesn't mean it's an ISP gaming speed tests - though perhaps they'd be more keen on peering directly and avoiding congested transit if it makes them look better, and I would absolutely believe that would happen.

[67726e]

I think the argument is that the speed tests are testing ideal/non real-world usage. That seems an awful lot like gaming the results. A more realistic, indicative speed test might hit several servers at points around the country, or even world. Most of my internet traffic is not hitting a server on the outskirts of town or in the next small city over.

[Vraxx]

Ugh. That's terrible. I kinda figured that it's an eventual conclusion though, so when speedtest.net starts giving me contradictory results to what I am experiencing I have a pretty good idea of what the check first.

On a separate note, what tool did you use to create that graph?

[icelancer]

Looks like just redirecting a recurring ping command output to a text file then graphing the results in google docs or any other spreadsheet. I've done the same to shame Comcast.

[cfallin]

Ooh, a chance to plug a toy side-project! I wrote "pingd" [1] to ping an address continuously and dump (timestamp, latency, dropped) tuples to a SQLite database, and also report results (tables only, no plots) via a web interface. Slightly less hacky than "ping host > logfile". Also to shame Comcast.

[1] https://github.com/cfallin/pingd (in Rust, requires liboping)