Re: Update of smoothed RTT in case of retransmitted segments

[Craig Partridge <[e-mail]>]

In message <1019653762.7923.30.camel@antares>, Sebastian Zimmermann writes:

>Hello,
>
>I need some help with the implementation of the smoothed RTT calculation
>in TCP in case of retransmitted segments:
>
>Assume that segment 9 is the one being timed. However, segment 8 is
>lost, so an acknowledgement for segment 9 cannot be generated by the
>receiver.
>The sender detects a timeout for segment 8 and retransmits it. Karn's
>algorithm only says that retransmitted segments may not be timed, but -
>in this case - segment 9 never was retransmitted, so the timer is not
>stopped. After the receiver receives the retransmitted segment 8, it
>will ack segment 9.
>
>The problem: the measured RTT is completely wrong, because it still
>contains the retransmission of segment 8.
>
>Since Karn's algorithm doesn't avoid this problem, which one does?

I haven't gone back to re-read our paper, but clearly Phil and I intended
that the ack be ignored, as it may implicitly contain the RTT feedback.

Craig