Zlatin Balevsky wrote:
The very idea of using a formula for making decisions about routing has one major flaw and that is the innacuracy of the estimators.
Well, classic routing routes on the basis of the assumption that getting data from a node means that it is a good place to get similar data in the future. That assumption can hardly be seen to be "accurate" in its own right - but it *becomes* accurate because of its own effect on the network topology. In effect, it is a self-fulfilling prophesy, and an extremely robust one at that.
We need to think about why NGR isn't benefiting from a similar self-reenforcing effect.
Given that NGR hasn't actually been implemented as intended, it's hard to say that it won't benefit from a similar self-reinforcing effect. There are two "proofs" for my assertion that NGR isn't yet implemented correctly: 1. the diagnostics indicate that the estimates are biased, 2. a derivation from first principles yields a very different formula. We should fix the formula and make sure that actual - predicted estimate() at least averages near zero (icing on the cake would be to make abs(actual - predicted) as small as possible).
Without at least that, we can't say that NGR has been implemented correctly, and we can't say what else we should do.
I've given more detail on what should be done to fix estimate() here: http://article.gmane.org/gmane.network.freenet.devel/8237 and here: http://article.gmane.org/gmane.network.freenet.devel/8143.
I don't think it is as simple as "its estimates aren't good enough", any more than it is valid to say that classic routing's routing decisions are perfect.
I agree with you there.
Ian.
I would like to get a conversation going on where we should go from here. I suggested an initial plan here: http://article.gmane.org/gmane.network.freenet.devel/8191. Later, I revised it here: http://article.gmane.org/gmane.network.freenet.devel/8236. So far, I haven't gotten any respone to the revised plan. What do you think about it?
-Martin
_______________________________________________ Devl mailing list [EMAIL PROTECTED] http://dodo.freenetproject.org/cgi-bin/mailman/listinfo/devl
