Router Fragmentation Is Gone from IPv6
In response to my Never-Ending Story of IP Fragmentation, Stojanco Cavdarov made an interesting observation: routers are not allowed to fragment IPv6 packets, they have to respond back with ICMP unreachable (effectively, routers behave as if IPv6 packets would have an implicit don't fragment bit).
To make life easier for non-TCP IPv6 applications (TCP is supposed to use Path MTU Discovery), the minimum IPv6 packet size that has to be supported on all links was increased to 1280 bytes (which, incidentally, fits very nicely into GRE+IPSec envelope transported across links with 1500-byte MTU).
I remember back in the mid to late 90's ISPs were very common to have link's MTU less than 1000 bytes (remember the famous Windows Dialup MTU tweaks?). Nowadays I believe most backhual links' MTUs are at least 4096 bytes within the ISPs themselves.
I wonder if the industry especially the router vendors should increase the default physical serial interface's MTU from the default of 1500 bytes to somthing like 2048 bytes going forward? I realize there could be potential memory issues for buffers, etc.; but for the time being 2048 bytes should be perfect to accommodate a single 1500-byte packet + IPSec + GRE header w/o fragmentation.
http://www.nil.com/ipcorner/IP_Fragmentation/
Stretch also wrote a great article about it:
http://wiki.nil.com/Path_MTU_Discovery