Reply to comment:

Geoff Huston
In the artlcle on large IPv6 packets I posed the question on whether it is preferable to use a 1,280 byte local MTU for IPv6, or use a 1,500 byte MTU (assuming that you are connecting to an underlying local network that can supports a 1,500 byte MTU, of course). We had experimented with a modified DNS name server that generated responses of 151, 1400 and 1700 bytes, and we looked at the loss rate per Ipv6 /64 at each of these sizes. Size Count Always Fetched Sometime Fetched Never Fetched 151 5,707 5,603 (98%) 104 (2%) 0 1,400 5,075 4,942 (97%) 95 (2%) 38 (1%) 1,700 5,086 3,857 (76%) 65 (1%) 1,164 (23%) The loss rate for packets of 1,400 bytes in size was 1%, which appears to be due to Path MTU issues for those end sites that site behind MTU-constrained tunnels. The far higher loss rate for 1,700 byte packets points to a set of problems with fragmented packets in IPv6 networks, including the dropping of trailing fragments and the silent dropping of packets that have Extension Headers. If the local host uses a lower MTU value, such as 1280, then it will cause fragmented UDP packets to start at the lower MTU size, and the not inconsiderable fragment filtering and Fragmentation Extension Header loss factors then come into play for our 1,400 byte test packet. This has now been measured in an experiment whose conditions are the same as the previous experiment with the single change that thge MTU size has been dropped from 1,500 bytes to 1,280 bytes. The loss rate for 1,400 byte UDP packets (which are now all fragmented) has risen from 1% to 20%, as shown in the following table. Size Count Always Fetched Sometime Fetched Never Fetched 151 4,211 4,067 (97%) 144 (3%) 0 1400 4,112 3,203 (78%) 75 (2%) 834 (20%) 1700 4,100 2,954 (72%) 84 (2%) 1,062 (26%) For the 1,400 byte packet, of these 834 failing IPv6 /64’s, 288 (35%) of these V6 prefixes generated a Fragmentation Reassembly ICMP error, indicating some form of packet firewall filtering system that is dropping trailing fragments. The remainder of the drops were silent drops. The message seems pretty clear that for UDP in IPv6 it's best for a sender to use a large MTU if they can, in order to avoid gratuitous fragmentation-caused packet drop. Geoff