1. You can compensate for most of the paths involving rate limiters by
the following trick: we have measurements in both directions of
each path. Eliminate any measurements to a destination that is a rate
limiter, instead using the reverse direction (from a rate limiter is ok!),
since all the measurements are round trip. As a result, we'd only have
to throw out paths between rate limiters, or roughly 1% of the total.
2. This suggests that there is a bug in the alternate path computation,
at least for drops. Amin ran a separate analysis of the data set
for his dissertation that pointed out the www.structured.net never got
better than 66% delivery to any peer. That should imply that there is
no alternate path to it that gets better delivery rates, but that is
inconsistent with the alternate path graph at the retreat,
tom