I suspect that more of these non-homogenous networks will crop up.
All the world isn't ethernet with a 1500 byte payload.
FDDI is (I think) 4500
4Mb token ring (about) 4472
16Mb token ring (about) 17800
Now with the larger sizes, filling the frame isn't as important, neither
are the above limits absolute - the token ring spec only defines a time
interval, the values above are from the source routing spec.

Other issues:
- The forward and reverse routes may be entirely different
  (eg phone uplink, satellite downlink)
- The routes might change during the lifetime of the connection

These all conspire to make anything other that the ultra-conservative
536 or something that dynamically detects network problems with large
fragments.

    David

> Bill Sommerfeld <sommerfeld@orchard.arlington.ma.us> writes:
> 
> > here's another picture:
> > 
> >____         ----           ____                ______        ______
> >|CL|----E----|NB|----POE----|RB|----Internet----| FW |--------| WS |
> >----         ----           ----                ------        ------
> > 
> > FW = "Firewall"; "WS" = "Web server".
> > 
> > "WS" sends out 1500-byte packets with DF set.
> > 
> > 1500 byte packet gets to "RB"; "RB" sends ICMP "fragmentation needed"
> > back to WS.
> > 
> > "FW", configurated in violation of RFC2979, blocks all ICMP packets,
> > including ICMP "fragmentation needed" messages, so WS never knows to
> > shrink its MTU to CL.
> 
> What Bill says. Here's another picture. "R" is a generic IPv4 router
> (e.g. Zebra/NetBSD) connecting one enterprise site to another over a
> private link. The "NB" routers are BGP peers to each other and to
> external ISP's.
> 
> Same problem. No PPPoE, just GIF wrapper for iBGP.
>               
> ____         ____  ___     ___  ____               ____      ____
> |CL|----E----|NB|--|R|-WAN-|R|--|NB|---Internet----|FW|------|WS|
> ----         ----  ---     ---  ----               ----      ----
>                \...  iBGP/gif .../