[ih] Arpanet raw messages, voice, and TCP

Noel Chiappa jnc at mercury.lcs.mit.edu
Thu Nov 26 11:53:24 PST 2009


    > From: Craig Partridge <craig at aland.bbn.com>

Ah, thanks for the in-fill. A few comments:

    > The problem is that you lose a fragment and all the other fragments
    > are now worthless.

The _theory_ was that if the packet were retransmitted exactly as-was,
then in a lossy environment, if the losses were stochastic (and as you
point out, often they weren't), you'd expect to often get the missing
fragment from the first instantiation out of whatever set of fragments
from the retransmitted instantiation made it through.

However, in addition to the problems you note below, many hosts could not/
did not retransmit an identical packet. Either they used a new IP ID
(since the tranport layer didn't _know_ the old one, if it was allocated
by the internetwork layer), or TCP created a new segment covering a
slightly different part of the sequence number space (common with TELNET,
since more input or output may have been buffered during the timeout
interval; some data may have been in an earlier packet and gotten ACK'd),
etc.

So then all the old fragments were useless (and took up buffer space in
the destination host, needed to be timed out there, yadda-yadda).

    > Floyd and Romanow showed that what I call the "dead fragments
    > walking" would then cause downstream loss of fragments of other
    > packets, causing more mayhem.

Oh, right, congestive collapse... We saw a lot of that, back in the day!

	Noel



More information about the Internet-history mailing list