I was heard to say a few days ago:
Please be aware that any traffic passing through FDDI XP's using DEC gigaswitches is biased. These switches break path mtu discovery.
This was in the context of measuring packet size in the Internet. After receiving a couple notes asking for clarification, I figured it was worthwhile to pass the same info back on to NANOG. Should one install a (fast|gig)ethernet card into one's Gigaswitch and said switch have FDDI interfaces, the switch will go from having a 4500+ byte MTU to a 1500 byte MTU. This is even between FDDI card to FDDI card on the same switch. The switch will happily fragment the packet if you leave DF off. However, if you turn it on, the switch will drop the packet silently. The switch is unable to generate ICMP and route it back to the sender. Given all of this, PMTU-D breaks in these circumstances. This is important because at least two XP's that I run route server service experienced problems due to this - Mae-West and PAIX. PAIX solved the problem by heavily encouraging customers to get off the Gigaswitch - I don't know if its still in service. Mae-West (SJ) still has this issue to the best of my knowledge. Mae-West (NASA) has been migrating to a SSR fast ether network to get off the Gigaswitches. Collectively, Mae-West still has this issue when transiting across the bridge. So, to make a long story short, try some PMTU probes on your switched fabric before trusting statistics about packet sizes. IM(probably undereducated)O, larger MTUs are not going to do you a lot of good across public exchange point fabric for a long time to come. Large MTU's across private interconnects, or specially targeted interconnects between "Service Providers" will probably be of greater benefit. -- Jeffrey Haas - Merit RSng project - jeffhaas@merit.edu