Re: TCP maximum segment size determination


Jeffrey Mogul (mogul@decwrl.dec.com)
16 Nov 1987 1726-PST (Monday)


PAP4@AI.AI.MIT.EDU ("Philip A. Prindeville") writes about
the use of an IP header option to probe for MTUs:

    What about the required overhead for gateways and routers to have
    to further inspect each packet?

(1) It should not be necessary to probe on every packet. All bets
are off, though, if the gateways are (a) oscillating, or (b) doing
loadsharing in a way that creates dramatically variable routes (a
mistake anyway, in my opinion).
(2) A well-designed probe protocol should not cost that much. Alas,
with IP we are stuck using IP header options, and they aren't that
cheap. If I were designing the next version of IP ...
(3) Source route already requires such inspection. (This is not
necessarily an argument in favor of a new IP option, but because of
points #1 and #2 the load should be a lot lower than source routing).

    It could be optimized so that only TCP packets are inspected, but
    still, that would seem to add to the burden of possibly
    compute-bound gateways...
Arggh! This sort of mechanism should NOT be TCP-specific ... that
makes the gateway logic more complex that it needs to be, it fails
to solve the equally pressing problem with other protocols (anyone
out there having problems with NFS/UDP fragmentation?), and besides,
it's not necessary ... because the source host is the one that decides
when to add the IP header option, it only needs to send the option when
necessary, and the gateways can tell quickly if any options are present.

A final point: if I thought that the best we could do with path-probing
is only "add only a little load to the gateways" then I would be against
it. The whole point is to
        add a little per-packet overhead
to avoid
        loading the network with retransmissions of lost packets
and hope that the latter outweighs the former.

-Jeff



This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:39:56 GMT