Buffer bloat is a problem. But, as he quite accurately puts it, there are no really good solutions out there. For those of us in the middle of the network, there is clearly no good answer. We need to provide RTT * BW * <some fraction>. We can argue about the fraction, but the RTT is a function of the flows involved anyway. As a flow can have an RTT anywhere from 0ms to 300+ms, we can't provide the _right_ answer.

The hosts can do better. As he points out, Linux is optimizing for long-haul performance, which is not entirely unreasonable. Linux _could_ do much better, but it would have to have buffering adaptive to the particular transport connection. Not there yet.

So in the end, he ends up throwing stones all over the place. Gee, thanks.