On Fri, Sep 21, 2012 at 07:39:31PM +0200, Florian Weimer wrote:
* Gijs van den Broek:
Hi Florian,
What's your stance on atomic fragments for IPv6?
We did not particularly consider atomic fragments. Could you be a bit more specific?
I think we've got a classical pick-any-two situation: interoperability with existing clients, statelessness in the server, and compliance with the existing specification:
You can send atomic fragments if you've recently received an ICMPv6 message requesting for a particular address. This requires state in the server.
You can unconditionally send atomic fragments. This breaks interoperability with existing clients.
You can never send atomic fragments. This is not what the specification requires.
This overlaps with concerns in your proposal because the size of the Fragmentation header might have some impact on your size calculations.
your still stuck guessing the PMTU. the problem is not just firewalls. the server is going to be keeping more state, either by this method or by switchig to TCP. i talked about this a few years back: ww.interlab.ait.ac.th/aintec09/program.php /bill