On Thu, Jun 07, 2007 at 11:23:48PM +0300, Jusa Saari wrote: > On Wed, 06 Jun 2007 19:11:27 +0100, Matthew Toseland wrote: > > > Recent probe data suggests a theory: > > > > Parts of the network are "rabbit holes" or "dungeons", i.e. sub-networks > > which are only weakly connected to the larger network. These cover a small > > chunk of the keyspace, say 0.36-0.41 (roughly, in the trace I had). A > > request for 0.5 got stuck down the rabbit hole. The gateway node at 0.436 > > was backed off; if the request had been able to reach that node, it would > > have been able to get much closer to where it should be. So the request > > bounced around in the dungeon, and eventually DNFed. > > > > What we need is some backtracking. At the moment backtracking only occurs > > when we actually run out of nodes in a pocket i.e. when it is really > > small. We track the best location seen so far on the request (not > > including dead ends i.e. RNFs), and whenever this improves we reset the > > HTL back to the maximum (10); otherwise it is decremented. > > No, what you need is an opennet. Having a "sparse" network with plenty of > "leaves" only connected to the rest of a network through a single node is > a natural feature of a darknet. The leaves are made of people who
Hi, the current growth model is not the only possible under a darknet approach. Yes, a problem is how the network grows. If no similar topology is in place from the start there is _no way_ the network (when scaling it) can become efficiently navigable by swapping positions trying to fit an ideal model of topology which does not exist (the Kleinberg model where it works well). One way that IMHO would grow the network better (assuming one wants to stay in the darknet approach) would be connecting growth to a (an existing?) 'killer app' that lets you easily invite your friends into the network. Currently the barrier is too high, and when wanting to get into the network you have no real incentive of getting friends connected compared to going to #freenet-refs I have heard of at least one proposal to make the network growth easier for friend-to-friend structures; there was some discussion on a jabber- plugin for easily letting you invite your friends. But that would only be a start, a nice one to evaluate though. > happened to be in #freenet-refs at the same time, exchanged refs, and > left; the connecting nodes are those who are running automated ref > exchange scripts and therefore get connected to the new leaves as they > are formed. Simply backing away from the leaves is going to overload the > single connecting node since all the traffic to and from the leaves is > going to go through it; and of course if that node happens to go offline > for any reason the network will splinter. Agreed, the load/congestion problem and the topology problems seem to be very related. Not just that under a poor topology much load may be placed on few 'good' nodes, but that we cannot expect the greedy routing to work very good at all as in the Kleinberg model (=> longer chains => more data for network). cheers, /v
