In message <200301101950.h0AJopR03169@s102-n054.tele2.cz>Jaromir Dolecek writes
>Another idea - possibly NetBSD side is waiting for the socket
>buffer to fill. Can you try if setting TCP_NODELAY option
>for the socket on NetBSD side would help?

Surely Nagle would only be triggered for the fragment at the end of
each 64k buffer write? If then?

Kyle: the very best approach is to obtain a libpcap packet trace
(preferably with tcpdump, on the NetBSD host), run Shawn Osterman's
tcptrace tool over the trace to produce `time-sequence graphs'.
Use   xplot to view the graphs. (Both tools are available via pkgsrc).

That should make any problems (and solutions) apparent -- that is, to
someone who understand TCP dynamics.