Since I've been threatening to do this for a while now...

Here's a very straight forward implementation of RFC3649 and 3742.

A quick summary is that it changes the TCP congestion avoidance algorithms
for Very Large congestion windows, by slowing down slowstart above a certain
threshold (net.inet.tcp.max_ssthresh, 100 MSS by default) and by increasing
the window quicker, and not cutting as much when in congestion avoidance.

This should help TCP on Very Fat pipes especially in combination with
SACK + NewReno.

Patch: http://www.sigusr1.org/~kurahone/tcp-highspeed-netbsd-2.99.11.diff.gz

-- 
Kentaro A. Kurahone
SIGUSR1 Research and Development