I built a graphical console-based speed tester (client+server) that you can run yourself to get a more accurate picture and test between your home and servers you actually use:
iPhone 6+ very close to router: 110Mbps 2012 MacMini (WiFi 802.11ac) very close to router: 120Mbps 2011 MacMini (ethernet but USB-powered switch in between): 92Mbps
With dslreports.com speed test the ethernet is consistently the fastest.
I have limited the download speed on the TP-Link Archer C7 router to 100Mbps in an effort to cut down on bufferbloat. Seems to work pretty well on the dslreports test - get A, A, A.
DOCSIS has some support for stream compression, so you'll often get highly unrealistic results.
On the other hand, basically all internet content is sent over the wire already compressed, so any stream compression doesn't help with it at all.
I've had speedtest.net overrate my connection by 2-3 times it's actual throughput.
Do you have a source for this? I've never heard of it and I can't find any reference for this or even proposed support for this.
That being said, when you know that Comcast and others prioritize their traffic through Speedtest AND you encounter subpar speeds even with this allocation in place, you know there is a problem.
Just use --listen-addr on the server side and on the client side, just run it like sparkyfish-cli <hostname:port>.
I think it's a good idea not to use standard ports, to avoid any sketchy prioritization practices on the part of the provider.
More generally, if your response to suspected bad network is to open speedtest.net, and given that a lot of bad network issues are transient, it's not at all surprising that you'll have cases where you suspect throttling, open speedtest.net, and the connection problem just happens to go away at that point. So this just becomes confirmation bias.
You'll have to set your own up on a server somewhere.