Re: A New TCP Retransmission Algorithm


Phil R. Karn (karn@flash.bellcore.com)
Thu, 9 Apr 87 13:00:47 est


You are quite correct in pointing out the oscillatory behavior in the
timeout interval when the network delay suddenly increases. I mentioned
this in my note, as I have both simulated the behavior you describe and
observed it in practice. The important thing, though, is that the
oscillation always damps out at the correct new value. How *fast* it
damps out depends of course on the smoothing function in use. I've been
using Dave Mills' nonlinear function. It seems to work quite well since
it responds more quickly to increases than to decreases in observed network
delay.

The "oscillation" is the result of a compromise between assuming the timeouts
are the result of sudden network congestion and assuming the timeouts
are caused by packets being lost outright because of link transmit errors.
Without information in the TCP header indicating which transmission
a given ACK is answering (clearly the ideal solution except for the
compatibility problems) my algorithm seems to be a workable heuristic.

Phil



This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:38:07 GMT