On Wed, 2003-11-19 at 04:56, Ian Clarke wrote: > Niklas Bergh wrote: > > NGR is designed to take into account a nodes available bw and include > > that fact in routing decisions. Instead of overloading an already > > overloaded (bw-wise) node even more NGR would be sending the Q to > > another node. > > > > A bw-caused QR will completely hide the actual bw problem from other > > nodes' NGRT:s since the 'load' that is actually causing a QR includes > > multiple other factors. > > I think I agree, NGR should handle bandwidth limitations, using QRs is > too crude a tool and as Niklas points out, may even hide the problem > from NGR forcing us to rely on the backoff mechanism. >
It seems to me that NGR can't possibly do certain things. When the standard deviation exceeds the mean, you can't even predict the sign of your random variable. NGR doesn't seem to be figuring variances of the estimators. If it did, sometimes they might be very large. Differences between estimators which are less than 1 s.d. are not statistically significant. Making decisions based on them is simply making random decisions. Worse than that, it may be making non-random choices between equally good alternatives, when a random choice is far better. In the case of bandwidth, suppose we have 10 nodes (call them servers) which are all reasonably good at the same region of keyspace. Then initially, NGR has no reason to favor one over the others, but insignificant differences in estimators will cause one to be favorite of the clients. This one server gets an excessive number of trailers going, and pretty soon, all of the clients decide it is no good at what it actually is good at. Then they go after one of the others, and pretty soon all of the clients have decided that none of the 10 servers is any good at what they are all actually good at. This happens because of failure to QR when too many trailers are active. Also, suppose that the servers go down once a day. If each server has a 25 hour backlog, then none of the trailers will finish, and all of the servers get a bad reputation. But if the servers accept a few requests and do a good job on those, then the estimators in different clients' routing tables will diverge depending on which node served the request, and which QR'ed it. This is exactly what we want. NGR should be arranging for 1/10 of the clients to use each of the 10 qualified servers. Also, NGR should be doing a statistical significance calculation, and when the difference between several servers is insignificant, it should choose randomly. This will cause different clients to favor different equally qualified servers, distributing the load. Comments? -- Ed Huff
signature.asc
Description: This is a digitally signed message part
_______________________________________________ Devl mailing list [EMAIL PROTECTED] http://dodo.freenetproject.org/cgi-bin/mailman/listinfo/devl
