congestion: avoid double-reducing CUBIC fast convergence#2640
Conversation
Quinn mistakenly applied the beta multiplier to the newly adjusted `W_max` instead of the current window. This double-reduced the window to 59.5% instead of the intended 70%. This fixes the math by snapshotting the original window for the multiplicative decrease.
I've tried to backport it |
Ralith
left a comment
Ralith
left a comment
There was a problem hiding this comment.
Great catch, thank you! Our congestion control performance has always been a little worse than it should; I wonder if this is why... any chance you have before/after data for some interesting case?
Unfortunately no :( I had issues with cubic previously with LFN network(500mbit, 200 ms latency) slowing connection down to several kib/s. These findings are just comparisons of sources with rfcs and quiche by llm. I can check other congestion controllers later |
3 participants
Quinn mistakenly applied the beta multiplier to the newly adjusted
W_maxinstead of the current window. This double-reduced the window to 59.5% instead of the intended 70%. This fixes the math by snapshotting the original window for the multiplicative decrease.