[Babel-users] Summary from Brussels meeting

[Juliusz Chroboczek]

Hi,

As many of you may know, Zoobab (Benjamin) organised a Babel meeting in
Brussels this last week-end.  I couldn't stay until the end (perhaps
Zoobab and Pieter can fill us in), but here are a summary of the part
I was present at.


First of all, under no circumstances should you stay at the Hotel Ibis
Sainte Catherine in Brussels.  Please see

  http://www.pps.jussieu.fr/~jch/ibis-sainte-catherine.html

(in French).


Second, Gabriel and I have been doing some serious testing of Babel-Z over
two hops with two radios.

     2.4 GHz
   A ------- B ------- C
     .......   .......
      5 GHz

We've found that it in this degenerate configuration, it's surprisingly
tricky to exhibit the Z effect:

  (i) when there is no packet loss at all, there is a 1/2 chance that
      plain Babel will pick the optimum route, thus being just as good as
      BabelZ.

      Indeed, there are four possible routes, of which two are optimal.
      In plain Babel, these routes have the same metric, hence 1 chance in
      2 to pick an optimal route.  BabelZ will of course pick one of the
      optimal routes.

      This doesn't happen with moderate packet loss, when the packet
      loss is different on the two frequencies, thus making plain Babel
      choose the bad route.

 (ii) when there is too much packet loss (more than 20% or so), there is
      very little interference between the different hops, and hence the
      optimal and suboptimal routes have almost the same throughput.
      Hence, our tests in that case show little difference between plain
      Babel and BabelZ.

(iii) Finally, when there is just a little packet loss, plain Babel
      consistently picks a sub-optimal route, while BabelZ consistently
      picks the optimal route.  In that case, we're witnessing TCP
      goodputs of 3.5 MB/s for BabelZ against 1.8 MB/s for plain Babel.

Our conclusions are as follows.  First, BabelZ works as expected, and
yields the expected performance -- almost doubling goodput over two
hops.  Second, the topology that we tested is too simple to give a good
test case, we really need to have four routers and three frequencies.

All in all, a productive week-end.  Thanks again to Zoobab and Pieter
for inviting us all.

-- Juliusz