Path: utzoo!attcan!utgpu!dennis
From: dennis@gpu.utcs.utoronto.ca (Dennis Ferguson)
Newsgroups: comp.protocols.tcp-ip
Subject: Re: ICMP Multiple Echo Replies
Keywords: ICMP echo reply
Message-ID: <1989Jul2.144711.7892@gpu.utcs.utoronto.ca>
Date: 2 Jul 89 18:47:11 GMT
References: <1131@umn-d-ub.D.UMN.EDU> <25908@agate.BERKELEY.EDU> <17125@bellcore.bellcore.com>
Reply-To: dennis@gpu.utcs.utoronto.ca (Dennis Ferguson)
Organization: Mechanical Engineering, University of Toronto
Lines: 32
Checksum: 38055

Without implying that it is necessarily a cause of duplicated
datagrams, I note that a receiver on a busy HDLC link can receive
duplicate copies of the same frame quite often if the error rate
is high and link-level retransmissions are being done.

The acknowlegement-retransmission scheme HDLC uses requires that the
transmitter resend not only a frame which was received in error, but
also all subsequent frames that were sent before the acknowledgement
indicating a frame was lost arrives.  If the transmitter is busy this
means that not only will the lost frame be retransmitted but almost
certainly the frame following the one in error and perhaps several
more as well.

Thus the receiver may often receive one or more frames following
an error twice.  According to the spec the receiver is supposed to
drop all frames until the errant frame is received successfully.  If
I were implementing this scheme, however, and knew the link was only
going to be used for IP and protocols equally unaffected by out of
order delivery, I would be very tempted to hedge my bets by sending
on everything that arrived without error regardless, this allowing me
to lie a little bit and keep the window advancing if an error occured
in the retransmission of a frame I got right the first time.

Then again, maybe not.  What I have observed is that when our link to
the NSFnet (30-some kbps, between Proteon routers) becomes loaded it
will produce an occasional duplicated packet.  I'm pretty sure this
isn't an ethernet problem since the frequency of occurance seems to
be intimately related to the load on the slow link.  I always wondered
why this happened.  The above seemed a plausible guess.

Dennis Ferguson
University of Toronto