On Sun, Dec 30, 2001 at 11:58:53PM -0600, Edgar Friendly wrote: > > Increasing maxNodeConnections I can go along with. It probably isn't > > necessary to mess with the maximumThreads setting (at least, we should > > fix the more obvious problems before making more subtle tweaks whose > > effectiveness will be difficult to assess). > > > Fair enough. I was just looking at the connection problem as a > slot/tab imbalance, and increasing the number of slots while > decreasing the number of tabs seems to be a good solution.
It is futile even to try such measures to deter transient nodes from loading the network, I think. We can try to limit their default speed of producing traffic but we can't control how many of them there are. > > But I was wondering why we don't just do away with maxNodeConnections > > entirely (after all, arbitrary constants are bad), and let the > > availability of threads be the sole determinant in whether a connection > > will be accepted. It is not a 30-second hack of changing a constant > > config value, but I believe it could be achieved fairly easily. > > > I just worry about a node recieving hella requests in a short time and > not being able to route out because it can't make enough outgoing > connections. I like having the incoming connections being a bit more > than outgoing, but still allowing for (what should be) plenty of > outgoing connections. The system I have in mind would have to give threads preferentially for routing and other duties before giving threads to accepting incoming connections, whenever threads are scarce. Then you naturally achieve the balance of incoming connections the node can handle within its thread limit. > > Also, I think we should greatly reduce the idle connection timeout. It > > is currently at 3 minutes. 10-15 seconds is sounding more reasonable to > > me at this point (my intuition is that it should not be significantly > > larger than the time it takes to set up a new connection). Remember the > > value was effectively zero in the 0.3 network, which, if my sense of > > history serves, had fewer problems with nodes rejecting connections. > > I agree that the idle timeout should be decreased. I disagree with > your analysis of the 0.3 network. The 0.3 network worked because > there was no pre-requisite for contacting a node (like knowing its key > is in 0.4), so you could contact nodes willy-nilly and you'd always > connect. Now there's the problem of nodes changing identities and a > host of other connection errors that can occur. I was just recalling that we didn't hold connections open in 0.3 (which was really before my time). -- :: tavin cole (tcole at espnow.com) :: _______________________________________________ Devl mailing list Devl at freenetproject.org http://lists.freenetproject.org/mailman/listinfo/devl
