On 04 Jan 2012, at 10:24 , Philip Homburg wrote: > Yes, but PMTU failures are not a protocol issue. > Is it is an operational issue. So when the IPv6 > network is just a bunch of techies who are connected by > tunnels, you expect PMTU to sort of work.
I didn't expect it to work, given the history of PMTU in the deployed Internet. When it does work, PMTU is wonderful. It has not been reliable for IPv4 either. > By the time the internet is big is enough that some > routers just send ICMPs with link local source and > nobody notices, then PMTU starts to break down. We have long standing experience with PMTU. It hasn't been completely reliable for IPv4. It is no surprise that it is not completely reliable for IPv6 either. > I'm sure you know about the small difference between IPv4 and IPv6 > when it comes to fragmentation. For IPv4, if a DNS server need > to send a, say, 1000 octet reply then it can just send it. > If necessary, routers on the way will fragment the reply. Actually, IPv4 routers generally won't fragment the reply for too-big IPv4 packets -- not even if the DF bit allows them to do so. Most deployed IPv4 transit routers disabled all router fragmentation of IPv4 packets years ago. My understanding is that existing DNS servers often choose to send less-than-MTU sized DNS replies in part because of issues with IPv4 PMTU. So, again, this situation is not new with IPv6. > On the other hand, for IPv6, a DNS server will have > to fragment at the lowest common denominator. So making > the minimum link MTU 576, will cause a lot more IPv6 > fragments then you would get for IPv4. And makes IPv6 > quite a bit worse than IPv4. That does not sound identical with other folks' analysis about DNS. For example, that isn't identical with what Mark Andrews has said. > If you follow this to the logical conclusion, then with the > IPv6-IPv4 translators, a DNS server has to add a fragmentation > header to every DNS reply, even the small ones. That does seem consistent with what other folks concerned about DNS have already noted. > Well, you can always tunnel IPv6 over IPv4. Problem solved :-) If there were infinite bandwidth, it would be. Sadly, RF links with smaller MTUs generally are also relatively low data rate -- mostly visibly lower data rate than Ethernet. Yours, Ran -------------------------------------------------------------------- IETF IPv6 working group mailing list ipv6@ietf.org Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6 --------------------------------------------------------------------
