On 3/31/2017 6:41 AM, Bless, Roland (TM) wrote: > Lars Eggert said something like: > "Bufferbloat is something that a congestion controller > can't really do anything about anyway... > it's not necessarily a congestion control problem." > > I believe that this statement is not correct. FWIW, I can see both sides.
I agree with Roland that a single endpoint *can* avoid the impact of bufferbloat *on its individual connection*, e.g., using delay-based feedback. However, having listened again to Lars's comments at the mic, I infer that he's right too - no amount of pairwise congestion control can avoid the aggregate effects of bufferbloat on a path. If others are creating the bloat - individually or by aggregation - then nothing you can do will avoid it. I.e., a solution to bufferbloat always requires fixing the queuing in the buffers that bloat. It's critical to ensure that flows don't create bloated delays for each other, but much less critical that they necessarily correct for the mistakes of a single endpoint pair (unless that's part of fixing the former). Joe
