Subject: RFC describing exponential backoff for TCP retransmission?
To: None <tech-net@NetBSD.ORG>
From: Kevin M. Lahey <firstname.lastname@example.org>
Date: 03/16/1998 18:25:00
I think I should know this, but I don't. Can anyone tell me what
RFC documents exponential backoff for TCP retransmissions?
RFC1122 seems to refer to an exponential backoff, but doesn't
really specify how it should be done.
Using NetBSD-current on the local LAN, I find that I get
12 retransmissions, each one second apart, and then the connection
gets dropped. The other OS I had close at hand, IRIX, does
exponential backoff, pausing 3 seconds, then 6, then 12, and so on.
Stevens suggests that systems ought to do this kind of exponential
Arguably, if the RTT is 12ms and the varience is 1ms, then
the RTO should be something like 16ms, and we're just rounding
it up to a second, and all is copacetic. I'm not sure I buy that,
and I'm not sure I want my system to work that way on the LAN.
I'd like to fix NetBSD so it works right, but I'd like to see
what "right" is supposed to be before I fix it.