Supernet block sizes - recommendations?

Thu May 20 20:45:03 UTC 1993

> > Usually these arguments do the trick-- its a matter of technical sense
> > prevailing over 'I want a class B' pride.
> 
> We ought to have a better response to the claim: "I want a class B to
> improve performance, (because I can easily make my UNIX software use larger
> MTUs)."

If this is critical, you can suggest modifying the constants in
/sys/netinet/in_proto.c (at least on Suns running SunOS 4.x):

	/*
	 * Default TCP Maximum Segment Size - 512 to be conservative,
	 * Higher for high-performance routers
	 */
	int     tcp_default_mss = 512;

Yes, bumping up tcp_default_mss will do so for all destinations (unless the
final destination requests a lower mss).  This is not without its problems:
it may have a detrimental effect on traffic passing through paths with a
lower MTU (causing fragmentation, which we know to be harmful), as well as
cause extra serialization delay on slowish serial links, causing slower
response time for simultaneously running interactive sessions.  The correct
fix is to implement MTU discovery, I guess, but not many vendors do that
yet...

- Havard