[e2e] RTT estimation
Craig Partridge
craig at aland.bbn.com
Wed Apr 30 12:41:50 PDT 2003
In message <1051727351.1782.94.camel at enpc3334>, Dong Zheng writes:
>My question is whether there is any metric that tells how well the
>current Jacobson/Karn algorithm doing the estimation? I mean, given the
>fluctuation of the RTT, to what level of the fluctuation can the current
>algorithm track?
I'll start by saying I'm not aware of one and then try to summarize a
conversation that Van Jacobson and I had on a hotel bench c. 1989 on
the topic.
First, the key issue is less the RTT than the RTO -- that is, when does
the round-trip timer go off. In 1989, when the emphasis was on maximizing
the value of every packet, the goal was to never have an RTO go off early
(where "early" is defined as -- the ACK is on its way but hasn't arrived yet),
and, subject to never being early, have the RTO go off as soon as possible
once it is clear that a packet is lost. In today's world where performance
is often valued more highly, you might seek a more aggressive RTO (i.e.
accept more early RTOs in return for better RTO times when packets are
lost) -- although I'll quickly note that the consequence of a timeout
can be resetting the TCP congestion window.
So, to my mind, if you want to be quantitative, you should assign a high
error weight to early RTO values, and a low error weight to slightly late
RTO values (but progressively higher, the later they are -- where "later"
means how long after an ACK should have come back but didn't, do we wait).
Craig
More information about the end2end-interest
mailing list