Re: TCP maximum segment size determination


Jeffrey Mogul (mogul@decwrl.dec.com)
16 Nov 1987 1714-PST (Monday)


JBVB@AI.AI.MIT.EDU ("James B. VanBokkelen") writes that
    If gateway gurus saw their way clear to do so, they might help some
    fraction of the world by arranging that IP fragments aren't transmitted
    consecutively (if there is other traffic to handle) or by inserting a
    little time delay if the Ether or other non-serial media is idle.

Alas, this doesn't work if there's another gateway further along the
route, since that gateway has no practical way of knowing that two
fragments arriving with a slight delay between them are fragments of
the same datagram. This gateway will normally "iron out" the delays
inserted previously, especially if there is any queueing delay (say
because the first fragment is huge and the second is tiny).

At Stanford we tried to play a similar game by fixing the gateway
(which was actually one of JNC's prototypes for the Proteon gateway)
between the ARPAnet and our ethernet to delay between any pair of
packets to the same local-wire destination. If all the gateways
did this, in theory it would work because then no matter what path
the fragments took through the Internet, they would always be spaced
out at the last step. Unfortunately, this didn't work 100% because
delaying for any reasonable period of time can't guarantee that a
timesharing system (with poor interrupt latency) will always have
time to get ready for the next packet, and because it can always
be distracted by local-net traffic (especially broadcasts and
multicasts).

-Jeff



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