Measuring Starlink TCP performance

In this episode of PING, APNIC’s Chief Scientist Geoff Huston discusses Starlink again, and the ability of modern TCP flow control algorithms to cope with the highly variant loss and delay seen over this satellite network. Geoff has been doing more measurements using starlink terminals in Australia and the USA, at different times of day exploring the system behaviour.



Starlink has broken new ground in Low Earth Orbit internet services. Unlike Geosynchronous satellite services which have a long delay but constant visibility of the satellite in stationary orbit above, Starlink requires the consumer to continuously re-select a new satellite as they move overhead in orbit. In fact, a new satellite has to be picked every 15 seconds. This means there's a high degree of variability in the behaviour of the link, both between signal quality to each satellite, and in the brief interval of loss ocurring at each satellite re-selection window.



Its a miracle TCP can survive, and in fact in the case of the newer BBR protocol thrive, and achieve remarkably high throughput, if the circumstances permit. This is because of the change from a slow start, fast backoff model used in Cubic and Reno to a much more aggressive link bandwidth estimation model, which continuously probes to see if there is more room to play in.