[This is in response to a note on tcpsat about retransmit timers
firing too early on long connections.]
I have seen unwanted retransmit timouts (i.e. bugs) of this nature in
two other implementations as well. All of these implementations have
in common more accurate timers. I suspect it would be useful to have
more detailed documentation on the use of timers and computation of
RTO. However, it isn't clear to me that there is a lot of consensus
on a number of the issues.
Here are some issues which I think are unresolved, or could use better
documentation:
- How to modify the RTO calculation when doing 1323 timestamps. The
idea here is that you are getting measurements more quickly, so you
might want to adjust the moving average calculation.
- Running with fine-grained clocks.
1. Should you?
2. Are adjustments to the computations needed? (Remember delayed
ack might tack on 200msec).
- Always time the RTO from the last packet in flight.
- Should there be a minimum RTO?
- Should there be a minimum deviation?
Interestingly, I see no discussion of these issues in the TCPSAT
drafts, and I don't believe I have seen them covered in any of the
other relevant places either (i.e. pilc or tcl-impl).
I'd be happy to pull together ideas on this subject if people are
interested in seeing such a document.
--Jamshid
PS. On a process note, I'm thinking this belongs under tcp-impl, so
I've sent this note to tcp-impl and BCC'd tcpsat (where the
conversation started).
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:55 EST