[Olsr-dev] Seeking comments: OLSR+ETX v/s DSR+ETX

Sven-Ola Tuecke (spam-protected)
Thu Dec 27 11:34:45 CET 2007


Hi Jens,

(not sure why everybody is complaining about the "fast-start" formula. This 
is the second wave <ggg> Yes- I'm not mathe crack - but to my impression 
that works. Hopefully this is not the "everyone talks about the 
fisheye-rhythm-table, but nobody notices for years, that fisheye simply does 
not do something" case)

A weighted packetloss formula will adapt faster to the changeing network 
conditions, right? Well - may be a good idea. We should get more realistic 
route adaptions. May be a bad one. We got too much route adaptions because 
minor and short-time effects will lead to wild and too-fast adaptions. 
Something like this needs verfication / falsification in real live IMO.

Personally, I tend to the latter. Because packetloss is ~half of the truth 
(e.g. influenced by link payload, appears to be bursty) and the current 
"spread the topology knowledge" may not be fast enough to cope with too 
frequent changes.

// Sven-Ola

"Jens Nachtigall" <(spam-protected)> schrieb im Newsbeitrag 
news:(spam-protected)
> -- 
> Olsr-dev mailing list
> (spam-protected)
> http://lists.olsr.org/mailman/listinfo/olsr-dev

> What's about this ? It's not as "fast" as the old formula, but I think it
> works:
>
> lq = (total - lost) * / (  (total + olsr_cnf->lq_wsize) / 2 );
>
> (just using the average of the maximum lq window size and the active one)
>
> This way fresh links can be used immediately, but get "bad" lq values
> until olsrd receives enough hellos.

not related to the startup, but afterwards:
What is still missing imho is a weighted, maybe even exponentially weighted
average (see formula at
http://en.wikipedia.org/wiki/Weighted_moving_average#Weighted_moving_average
and below). Pakets just immediately lost should be paid more attention than
towards packets that were lost one or two minutes ago. For instance, in the
FreifunkFirmware HelloInterval is set to 5 seconds and LinkQualityWinSize is
set to 100 samples -- hence the time window is 500sec, that is more than 8
minutes. Still, to the formula it does not make a difference if the packet 
is
lost 8 minutes ago or just few secs.

regards,
Jens






More information about the Olsr-dev mailing list