> But this is not what RFC1122, Page 57, Paragraph 3.3.3 states. It
> specifically states that exceeding the MTU is not allowed and that
> a mechanism for determining the MTU must be provided.
I think you are confused. That is correct for a host that does not do
local fragmentation, if you are talking about TCP/UDP. If you are talking
about sending packets larger than the accepted MTU on your locally
attached network, this is obviously true, but irrelevant.
IPv6 has a mechanism for path mtu discovery at the IP level. This was
introduced so that any gateway along the path will not have to do any more
fragmentation ... because it is memory and cpu consuming. This is the
general ideea in RFC1122, that for performance reasons, you should choose
a path MTU for minimum fragmentation/reassembly.
> What caught my attention was that someone was bothered that UDP would
> not handle a packet of many thousands of bytes (octets). I stated that
> it should not be used for this purpose.
NFS currently does exactly that ... (ok, you could argue that nfs is
broken anyway :))
> Further, whether or not the IP layer will fragment and defragment is
> moot. This is just a side-effect of IP and should not be used as a
> the basis of a communications design.
It is not really a side effect. IP fragmentation/reassembly was
designed specifically for this purpose, to allow networks with large MTUs
to comunicate with the ones with smaller MTUs.
> Cheers,
> DJ
> Richard B. Johnson
Matei