> I should probably mention I'm setting TCP_NODELAY to one on this, so I'd
> expect any such heuristics to be deactivated. :-)
> 
> Also, setting 'sysctl -w net.inet.tcp.ack_on_push=1' makes the problem go
> away, but at an increased packet flow.

Mmmm... the 'interwork with M$ stack' feature :-)

Obviously a timestamped line trace would be useful.

Still guessing....
If the 'ack_on_push' flag makes a difference, then clearly (?)
something is waiting for the data to be acked.
I wonder if the 'zero copy tcp' stuff is getting in the way
- somehow stalling until TCP releases the 'borrowed' buffer.
4k is (probably) the system page size....

	David

-- 
David Laight: david@l8s.co.uk