broken bandwidth-delay product window limiting

[David G. Lawrence]

> On Fri, Mar 25, 2005 at 07:29:10AM -0800, David G. Lawrence wrote:
> >    The "bandwidth-delay product window limiting" code in 5.x appears to be
> > quite broken. In my tests here, the window is hard limited to 2 segments,
> > regardless of the actual bandwidth-delay product, and this results in very
> > poor performance over non-local TCP connections.  I'd suggest shipping 5.4
> > and future releases with net.inet.tcp.inflight.enable=0 until this problem
> > can be fixed.
> 
> Have you seen:
> 
> 	http://www.FreeBSD.org/cgi/query-pr.cgi?pr=75122

   I have now. :-)
   I saw parts of the discussion when it was occuring, but didn't really
put it all together.

> I wonder if it is the same bug or a different one?

   You know, it's really hard to tell. This code is very complicated. All
I've been able to determine is that many tests with a variety of non-local,
but 100-1000Mbps connected machines, all resulted in a 2 segment limit on
the TCP window with top performance of less than 1.5Mbps. With the old
not-so-clever algorithm (inflight.enable=0), I see about 15 times this with
sendspace=64K.

-DG

David G. Lawrence
President
Download Technologies, Inc. - http://www.downloadtech.com - (866) 399 8500
TeraSolutions, Inc. - http://www.terasolutions.com - (888) 346 7175
The FreeBSD Project - http://www.freebsd.org
Pave the road of life with opportunities.