Hi Stef, I forwarded your message to devl.
Am Sonntag, 6. Juli 2014, 23:01:00 schrieb Stef: > from looking over the logs (I hope I found all the points concerning us ;) ), > one main point of discussion is the simulation study about the impact of a > suboptimal distance distribution on the routing. > 1) The study is, as already noted in some of the logs, not really realistic, > because we use neither caching nor churn nor FOAF-routing. I think we missed that. > It only shows that the routing performance in Kleinberg's model (so an > artificial network model) is drastically decreased if long-range neighbors > are chosen uniformly at random rather than proportional to 1/d (d = > distance). We changed Kleinberg's model slightly to allow for arbitrary > degree distributions and used the measured degree distribution in Freenet. An > implementation can be found here: > https://github.com/stef-roos/GTNA/blob/grouting/src/gtna/networks/model/smallWorld/KleinbergDegreeDist.java That’s cool - thanks! > So the actual hop counts are likely very different in the real network (so > best forget about those numbers ;)). However, it seems reasonable that the > routing in actual Freenet Opennet is worse than it could be as well. (caching > might mitigate the effect to some extent...) We could find other indicators for that misrouting. For example the distance of the originating peer of a request decreases for the first hops (as it should be) but then actually inreases at low HTL. This suggests that the requests might bounce around on random routes. The much higher success-rate in realtime routing suggests, that part of our problems could also be due to overload (realtime routing has different bandwidth-limiting which is optimized for latency instead of throughput, and that means that it operates with less connection overload and consequently with less misrouting. Our stats suggests that we only use 15% of the connections for short-distance routing, while in a Kleinberg-model that should be 80%. Consequently moderate overload could block the short-distance routing completely for popular paths, essentially reverting freenet to random routing. > 2) differences to Oscar's simulation: Which simulation are you referring to? > I assume those in Distributed Routing in Small-World networks, where he > compared random ID placements for Darknets with swapping? well, he used a > rather high uniform degree (6 log_2 n are more than 80 neighbors per node) > for all nodes while we used a non-uniform degree distribution with a lot of > nodes with a degree of less than 10, that will lead to different results, > especially if the assignment is random and the target is found by chance > because it is the neighbor of a contacted node... There are different simulations, some also from our own simulator. I’m not sure which one of these was meant, but I hope someone on devl can answer this. (toad?) > 3) yeah, binning and restricting connections, not exactly an elegant > solution, but it seems like you couldn't think of anything better either... > I considered doing some strange statistical tests for checking locally if > the neighbor selection could be generated by a Kleinberg distribution, but > the locally available samples are probably too small for significant results. > Even if they were, a test saying that the distribution is not good does not > tell you what to do *sigh* The main complication is that we can only use local information, and it is possible to construct networks which locally look like a nicely routable small-world network but which have a broken global structure. That’s why we use the only trustworthy global information we have: successfully retrieved data (chunks). Here’s a sketch of what we’ll test: https://d6.gnutella2.info/freenet/USK@ZLwcSLwqpM1527Tw1YmnSiXgzITU0neHQ11Cyl0iLmk,f6FLo3TvsEijIcJq-X3BTjjtm0ErVZwAPO7AUd9V7lY,AQACAAE/fix-link-length/3/ https://127.0.0.1:8888/USK@ZLwcSLwqpM1527Tw1YmnSiXgzITU0neHQ11Cyl0iLmk,f6FLo3TvsEijIcJq-X3BTjjtm0ErVZwAPO7AUd9V7lY,AQACAAE/fix-link-length/3/ > anyway, cool that you figured out why the distance distribution is that > strange, we tried by looking at the code, but didn't find the reason there ;) We’ve been discussing about that topic for a few years, now, so the ideas for the reasons are not completely new. Your clear writeup in your paper helped us get out of the block and have a focussed discussions on the link length instead of getting sidetracked by other related issues which could also be a cause for misrouting. http://127.0.0.1:8888/USK@s9sxY2cTJWHKRsTuBTkjrXW4HfzrdUlwFqft1mzV0Gs,2E4DOMYy-~zOdp8-5OQH2IcmLfey0AOIkms-73Mx2tI,AQACAAE/freenet-funding/11/ > > https://d6.gnutella2.info/freenet/USK@s9sxY2cTJWHKRsTuBTkjrXW4HfzrdUlwFqft1mzV0Gs,2E4DOMYy-~zOdp8-5OQH2IcmLfey0AOIkms-73Mx2tI,AQACAAE/freenet-funding/11/ > ah je funding, what type of project are you applying for exactly? If it is > the type where a university is useful as a partner, we would certainly > consider it (I assume, that is the kind of decisions I can't make given that > my time at TU Dresden is limited) I think a university as partner would be a big advantage: While we want to make Freenet useful for practical applications, which will first of all require lots of work on non-structural problems¹, there are still many academic challenges like checking which mitigation stragegies against the pitch-black-attack actually work without breaking routing. Best wishes, Arne ¹: non-structural problems like making it easy to give freenet to friends and have them connected via darknet right away without ever having to connect to freenetproject.org
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ Devl mailing list Devl@freenetproject.org https://emu.freenetproject.org/cgi-bin/mailman/listinfo/devl
