On 24-jul-2005, at 2:17, Bob Hinden wrote:
Don't forget that management issues in layer 2 networks (= a single layer 3 subnet) are very different from those on the global internet: most layer 2 networks are managed by a single organization. Since we had the hubris to think we could do PMTUD for the global internet, I think we shouldn't shy away from doing the same thing for layer 2. As always, the operator community will decide wheter what we come up with is usable in practice.
Fair point. As suggested on the list a new NA/NS option seems like a reasonable approach except for the problems pointed out. Having to do probes of different packet sizes is fairly messy if there is a device in the middle with a smaller MTU. Would a NA/NS option solution solve enough of the problem to make doing it worthwhile?
If in a conservative approach we don't want to send "jumboprobes" we still need an easy way to let the hosts know the limits of the infrastructure. A packet that is multicast by systems "in the know" would do the trick. Since the most obvious source of knowledge would be switches, it makes sense to make it easy for a switch to multicast this information, which means it shouldn't be part of RAs as switches as a rule don't send those out. (If we make the message very simple and let the switch use the unspecified address as a source address, there wouldn't be any need for the switch to run IPv6, it would just transmit a fixed content packet.)
In the absense of "smart" switches, one or more routers or even hosts could be configured to send out proxy jumboframe size announcements.
When a host hears multiple jumboframe sizes announced, it would select the smallest it hears.
To avoid the situation where there is a smart switch that announces a jumboframe size and also a stupid switche that doesn't, and also don't support jumboframes (of that size), or the situation where an administrator makes a mistake, it would still be necessary to confirm that the advertised MTU size is correct by sending an MTU sized jumboframe after an ND option indicates that a neighbor supports jumboframes.
This approach still provides most of what we'd like, except that all systems share the same effective MTU, which they learn from an administratively configured value, and hopefully in the future, directly from the layer 2 infrastructure.
What do you think? -------------------------------------------------------------------- IETF IPv6 working group mailing list ipv6@ietf.org Administrative Requests: https://www1.ietf.org/mailman/listinfo/ipv6 --------------------------------------------------------------------
