On Thu, Jul 17, 2003 at 10:40:37AM -0400, Zlatin Balevsky wrote: > if you plan on publishing a paper or something on NGRouting it may be a > good idea to investigate further into what kind of behavior does the > graph keyspace vs. time evolves. I'm very interested in studying which > method of estimating response time is best mathematically because it may > unveil some unknown facts about freenet routing in general.
Well - while this would be interesting, I don't want to make it a prerequisite for the article - as if I did then there would be 1001 other things that are similarly deserving of investigation and the article would never get written. Personally I suspect that the difference between the current simple linear interpolation mechanism and a more sophisticated curve-fitting approach will probably be minor at best - but the nice thing is that we will be able to test this easily once NG is up and running in the real world. > This brings us back to the phone discussion we had some months ago about > how to collect reasonable data. This time around we'll need a different > set of measurements, and since one of them will be exact sampling data > of requested keys we cannot assure any privacy for the testers. So, > could we adapt the watchme functionality to report that and enable it by > default in experimental? If NGRouting is hyped enough at the right > places we'll find plenty of people willing to test it. The problems of > the watchme network (i.e. unrealistic usage patterns) do not really > affect us this time, as all we're measuring is key response times. Again, one of the great benefits of NG routing is that we will be able to test its effectiveness based on local data only (ie. how close our estimates are to the actual routing times) which should reduce the need for some form of elaborate global data collection mechanism. > My guess/working hypothesis: during the initial stage while the node is > getting to know the network, we'll be better off interpolating a > periodic function; for heavily specialized nodes, even if they have > multiple specializations we'd be more efficient to use one (or perhaps > two or more independent) polynomial interpolations. Possibly. One important point to accelerate a node's ability to learn about the network is that when a node sees a reference (perhaps in the DataSource field of a DataReply) and it has statistical information (a RTE + other data) about that node, then it adds it to the reference, possibly replacing statistical data which may have been there before. This means that from the moment a node learns about a new node in the network, it should already have a good idea of its characteristics - rather than spending hours and hundreds of requests probing it itself. We should also consider having such information added to seednodes.ref files so that when nodes first "come up", they already have good network knowledge (although this may be dangerous in a situation where everyone gets refs from hawk and thus all end up with hawk's specialization to begin with). Ian. -- Ian Clarke [EMAIL PROTECTED] Coordinator, The Freenet Project http://freenetproject.org/ Founder, Locutus http://locut.us/ Personal Homepage http://locut.us/ian/
pgp00000.pgp
Description: PGP signature
