On Sat, Dec 26, 2020 at 7:28 PM Mikael Abrahamsson <swmike@swm.pp.se> wrote:
On Sat, 26 Dec 2020, Baldur Norddahl wrote:
I demonstrated that it is about buffers by showing the same download from a server that paces the traffic indeed gets the full 930 Mbps with exactly the same settings, including starting window size, and the same path (Copenhagen to Stockholm).
You demonstrated that it's about which TCP algorithm they use, probably.
All (virtual) machines used in the experiment are the same. Those are NLNOG RING network managed machines all running the exact same Ubuntu 16.04.7 LTS. If you have access to NLNOG RING or equivalent you should try the experiment for yourself. You will find that as latency increases TCP speeds goes down and it can not be explained by congestion. And you will find that some servers have this effect much less than others and that those servers usually have 1G network speed. The effect is the same no matter what time of day you try it (ie. it is not congestion related). Before you panic I will say I am not trying to advocate that we need more buffers. We need smart buffers. Buffer bloat is bad but no buffers is also bad. Your home made debloat solution will probably not be able to recover the missing TCP performance that I am describing here. But if you could have FQ Codel in the ISP switch that would probably do a lot. Or we could have TCP with pacing and that will be widely deployed around the same time as IPv6. Regards Baldur