TCP not following Jacobson rule
- To: email@example.com
- Subject: TCP not following Jacobson rule
- From: firstname.lastname@example.org (k1io, FN42jk)
- Date: Wed, 5 Feb 1992 12:08:05 -2359
Like dawn breaking over Marblehead, it dawned on me last night why
our local TCP LANs are such a mess (radio "LANs"). Watching the traces,
I never ever saw the TCP Window adjust. It always stayed at the fixed
level for that node.
That's NOT how TCP works. The host requirements specify the use of
the Van Jacobson backoff, which is essentially a dynamic window. If
a packet is lost, the retransmission should be sent with a window set
to 1*MSS, NOT the normal level. After that's received, it goes to 2*MSS,
then 3*MSS, then 4*MSS, until it hits half the value which it had hit
before it collapsed. Then it rises even more slowly. Note that the
increase must be _additive_, NOT multiplicative.
The details are in some RFC or other. It leads to stable networks.
In the NOS versions I'm seeing on the air, everybody sits at 864 or 432
and full-window retransmissions do often seem to occur.
Am I mis-reading the traces, or is there really a bug in NOS TCP?