[net2-wg] CTP experiments

[Philip Levis]

I've started experimenting with CTP's link estimation and route  
selection in TOSSIM using different link settings. My initial  
conclusion is that there's a relationship between the data estimate  
window size, the switch threshold, and alpha. Specifically, it's  
important that an isolated packet loss not cause a switch. If we look  
at the current settings:

switch threshold = 1.5 (you switch iff new candidate has a route ETX  
1.5 lower than current parent)
alpha = 0.2: new = ((old * alpha) + (1 - alpha) * new)
data window = 5

This means a single lost packet will give a data estimate of 1.25  
(5/4). If you crank this into the alpha, it causes an estimate of 1.0  
to go up to ((0.2) + 0.8 * 1.25) or 1.2.  If, with these settings,  
you had the switch threshold 0.1, then a single packet loss could  
cause a routing change, which is probably not a good idea.

 From a few experiments, it became pretty clear that the current  
switch threshold is *way* too high. I think that to start to really  
nail down what these values should be, I need to run a long series of  
simulation experiments with different values, which we can then  
compare with real network results.

For now, though, I found that changing the parameters to

switch threshold = 0.2
alpha = 0.5
data window = 5

was much more effective.

The simulations I ran used the current TOSSIM in CVS, which has CPM  
interference modeling (based on the meyer busy trace), handles ack  
losses like packet losses (uses an SNR curve), and has a 0.1% chance  
of a acknowledgment false positive.

Phil