On Wed, 21 Jan 2009, David Conrad wrote:
Paul,
I assumed there was an implied ", to a significant enough extent for it to
matter" at the end of the sentence I quoted.
What meets your requirement of enough "to matter"? Depending on
deployment scenario, a tunnel-router-based solution could cover
thousands of hosts or more. Is that enough to matter?
It's going to be subjective to a point, but determining "to matter"
would involve comparing the control-overhead to the actual
data-bearing traffic, rather than focusing solely on the difference
in control traffic between end-host and intermediary based solutions.
Having a tunnel-router do the liveliness tests for 1000 hosts might
save 999 instances of control traffic, but if the 999 instances of
control traffic comprise 1-2% of the data traffic - then should we
care? (Is there a data-communications version of Amdahl's Law I can
quote here?).
By my back-of-the-envelope calculations, a Shim6/REAP solution has
Order(1.2%) overhead relative to data-traffic (on a low amount of
traffic, 24kB - e.g. fetching the google front-page) for the normal
case. Failure cases presumably might take slightly more - but failure
is not the common case. Solid quantitative studies would be really
useful.
So basically, where's the awful overhead of host-based signalling,
relative to what matters: the actual data?
It seems it would be good to have data to back up such an assertion,
rather than accepting it blindly and introducing a lot of extra
complexity (and losing much fate-sharing) to optimise an apparently
relatively-trivial overhead.
regards,
--
Paul Jakma [email protected] [email protected] Key ID: 64A2FF6A
Fortune:
"I always avoid prophesying beforehand because it is much better
to prophesy after the event has already taken place. " - Winston
Churchill
_______________________________________________
rrg mailing list
[email protected]
http://www.irtf.org/mailman/listinfo/rrg
