Hi,
I tried looking at the cum dly vs throughput as Alistair suggested but
it's kind of weird and I think it actually takes more discussion. But,
we were talking about the drop
share not always being straightforward. I made a plot of the throughput
(of the "wifi"
example) vs time with the cumulative dropped kbytes. Note that tail drop
has the most
drops. Of course, this comes because of all the drops in the early stages.

Oh, black here is tail drop, green is codel, red is RED. Codel is a
little worse here than
the figure in my slide deck because things vary a bit with random seed.

I also plotted the number of drps for each, subtracting off the number
of drops each one
had at 4 seconds. Here you can see for most of the simulation RED has
fewer drops
but the throughput is not so great. Tail drop has the most drops and the
most throughput.
That transition you see just after 200 for Codel is because of the BW
drop and and that
it is adjusting the queue size (you can look at the plots in the slide
deck). I think there
are two take aways: 1) the relationship between drop rate and throughput
(as well as
delay) is not at all predictable and 2) Tail drop and RED are largely
dropping "the wrong
stuff"