[Olsr-dev] [OLSR-dev] OLSR for Windows Mobile platforms by?Moviquity

[sebastian sauer]

Hi,

Wed 15 Jul 2009 15:22, Henning Rogge wrote:
> Am Wed July 15 2009 15:06:31 schrieb sebastian sauer:
> > in general it's not trival to find a suitable metric for a highly dynamic
> > "loss network" like our mesh nets.
> ETX is a very "primitive" sollution for this and there might be better 
> metrics, but it's easy to implement and depends only on layer 3.
yep. and all "better metrics" I know build upon the the ETX. and not all
"better metrics" can guarantee "monotonicity".

ETX certainly leaves a lot to be desired. Yet even an real-life ETX
implementation is hard-core slavery / work.

I doubt any other OLSR implementation, has such a sound one as olsrd.
Extremely nice work what you've done here :)

> > hop count and ETX has some nice properties like "monotonicity" (some papers
> > call that "isotonicity" -- dunno why.)
> Isotonicity is a little bit different I think.
why do you think so? :)

Basically -- depending on the defs -- it's either the same, or slightly
different but it doesn't really matter. Both allow guarantee monotonic
decision logic.

Which is great because "non-monotonic (decision) logic" is quite a
mind-fuck.

if you like a (shitty) intro ;)
http://plato.stanford.edu/entries/logic-nonmonotonic/


> Isotonicity means that if a path between two nodes A,B is the shortest path by 
> a certain routing metric, any shortest path between nodes on the patch A,B 
> will lay on the path between A and B. This is a very important assumption for 
> hop-by-hop routing.
I aggree it's a "very important assumption" esp. for "hop-by-hop
routing". basically it's also an insurance against routing loop
disasters.

yet, some say Isotonicity is when the direction of the trend (in our case: the path 
ETX sum) is strictly increasing, while monotonic could imply a trend that is either
strictly increasing or strictly decreasing.

so from this POV we have monotonic anyway ;)

Srsly. Isotonicity, monotonicity are (almost) the same, important is that
they mean "order-preserving" in graphs like our mesh nets.

> > in practice things are even thougher -- none of the bigger OLSR mesh
> > setups work without ETX cheating. Hence on guarantee that the metric of
> > any randomly chosen path is really monotonic -- in a real life setup.
> What do you mean with ETX "cheating" ?
well this LQ factor / multiplier, what else? ;)

> > yep, hop count is an idiotic choice for a dynamic "loss network". because
> > the bandwidth of the "edges" / links is not stable over time.
> Let's say it was a decission done in a time where noone had experience with 
> large scale networks outside the simulator.
I disagree, there's tons of prior knowledge from "control theory" or
"complexity theory" only it was ignored.

e.g. take a so-called "crossing network" as a random example. Not really
different from an OLSR mesh. you have nodes, and edges with weighted
lables that represent transaction costs ...

*and* you have a significat possiblity for human crowd behavior, just
like in our OLSR mesh with LQ factor / multiplier ...