[SensorNetArch] A Thought on Routing + Congestion Control

[Cheng Tien Ee]

Here's an issue that has been growing on my mind in the past 2 weeks: 
congestion control typically requires some sort of push-back / 
re-feedback (this is a cool term) mechanism. What's more subtle is the 
fact that it also requires a relatively stable routing layer to be 
effective. Compare with TCP: if the routing is wildly unstable then the 
(implicit) feedback (such as RTT and occurrence of pkt drops) will be 
inaccurate, and TCP's performance becomes less predictable and will 
almost certainly deteriorate. In sensor networks, it seems that the more 
the routing fluctuates, the less relevant the feedback becomes, or 
another way of putting it is that convergence time increases. So the 
question is: how much should we damp routing changes: if we damp it too 
much, it helps cc indirectly because feedback then becomes useful, but 
we may be routing along links that have dropped in quality and thus we 
end up losing more packets anyway. If we damp less, we can be choosing 
better paths, but cc doesn't perform as well and we also end up having 
more drops. Thus if we plot the graph of packet drops vs. amount of 
damping we'll get a graph with a valley in the middle, or a trough, 
where packet drops is minimized. So the question becomes: is it possible 
to operate around that valley point?