[Tinyos-help] LinkEstimator, link qualities, ETX

[Philip Levis]

On Sep 1, 2008, at 4:56 AM, Michiel Konstapel wrote:

> Thanks for your input! Great to get feedback "from the source", as it
> were. Warning: wall of text coming up.


<snip>

I agree that the initial link quality should be 10 not 0; for some  
reason I think this fix is in the head?

What we tend to see over time is that CTP settles on very good and  
stable links. A link that *doesn't* have bursts of losses will hold  
longer than one that does. This has advantages in that route stability  
leads to fewer loops and control traffic for route repair. The cost  
(transmissions/deliveries) continues to decrease even after hours,  
seemingly approaching an asymptotic value.

The issue with a 1.5 threshold has come up before; Stephen at Berkeley  
noted that this means a node locking onto a 2-hop, ETX=2 route will  
never move back to a 1-hop, ETX=1 route. This assumes that the ETX=2  
route never varies (unlikely, at the very least due to collisions and  
interference). If you imagine both routes do not have perfectly  
stable, but rather time-varying ETX values due to dynamics, then at  
some point you'll have ETX=1 and ETX=2.5, at which point the node will  
switch back. The probability that you'll get ETX=3.5 and ETX=2 is much  
lower than ETX=1 and ETX=2.5, so the node will spend much more time on  
the overall better route. This handwaving assumes a bunch of things,  
of course. The best thing to do is to examine, through traces, how  
often this edge case occurs and whether it matters. What's nice is  
that we've reached a level of instrumentation and data gathering that  
we can really begin to quantitatively evaluate subtle changes.

Phil