On Sun, Nov 30, 2003 at 10:09:40AM -0800, Martin Stone Davis wrote: > Ian Clarke wrote: > > >Ok, the new stable build seems to be working quite well, are other > >people experiencing the same thing? > > > >We need to take stock of the situation with NGR. I think one problem > >has been a willingness to dream up solutions, and implement them, before > >actually understanding what the problem is. > > > >I would like to propose that now that the time-pressure is off, we try > >to be more cautious - we need to form theories about what the problem > >is, figure out how to test these theories, and if they prove true, > >*then* we implement a solution. > > Well, okay, but how does that relate to the plan I outlined > (http://article.gmane.org/gmane.network.freenet.devel/8191)? > > I understand why you would want more proof that we should drop HTL in > favor of time-to-live (Toad's idea, which I support). To do so *right > now* would be a big change, since we still have many details to be > filled in (http://article.gmane.org/gmane.network.freenet.devel/8184). > Therefore we should only do it if we're confident that it will cure a > major ill.
Unfortunately, we can't. Well, we can't easily. Because there will be a nonlinear relationship between pDNF and the time to live value. One possible solution is to use something more sophisticated than an estimator (some sort of machine learning algo maybe). Another solution would be to not have a variable timeout. Requests die when they go to all *available* nodes (which is limited under load by unobtanium [ only accepting requests we have a chance of completing ])... > > However, I think we can implement the other ideas right away. We know > (Toad has proven) that we have more queries than we can actually deal > with, so your back-off scheme should be implemented. We know (I have > proven) that the current estimate() formula is incorrect, so we should > fix it to match its original purpose. And as for Unobtanium routing, I > can't prove to you that it would cure a major ill, but it would not be > hard to implement, it couldn't hurt, and it just might help. We should think about tRetry. Maybe it needs to be infinite. I.e. maybe we need to route strictly by success probability. And if we don't, we have to somehow get rid of the HTL, because that's the only way not to have tRetry (failures such as connection failure could lead to a restart - what we'd get rid of is not just DNFs, it's all the varied reasons to kill a request). > > > > >One thing that is important is simply to figure out how accurate NGR's > >estimates actually are, and whether their estimates are statistically > >significant. Sure. How? > > > >Also, understanding which parts of the NGR estimate > >calculation have the most bearing on the routing decision. > > > >Thoughts? > > > >Ian. > > -Martin -- Matthew J Toseland - [EMAIL PROTECTED] Freenet Project Official Codemonkey - http://freenetproject.org/ ICTHUS - Nothing is impossible. Our Boss says so.
signature.asc
Description: Digital signature
_______________________________________________ Devl mailing list [EMAIL PROTECTED] http://dodo.freenetproject.org/cgi-bin/mailman/listinfo/devl
