On Mon, Sep 01, 2003 at 10:34:19PM -0700, pineapple wrote: > Aesthetics aside, I don't see how this can make the > problem worse. When the node runs out of threads it > is practically dead (upstream bandwidth is typically a > few hundred bytes/sec or less). Some connections get > killed and the node is able to carry on and be > usefull. Othwise, this condition can literally > persist for HOURS (depends on how busy the node is). > This then forces the network to route around the node > which hurts performance for everyone (possibly causing > DNF's). The test is simple, if upstream bandwidth is > very low and there are alot of > freenet.node.states.data.DataStateInitiator threads, > then kill some of said threads, starting with the ones > idle the longest. If none are idle, kill some > freenet.node.states.data.DataStateInitiator threads > with the lowest transfer rates. I understand that > this is an ugly hack, but I don't understand how it > could make things worse since it would only happen > when the node is basically dead.
Because killing transfers forces the requests to restart. Because handling new requests is LESS IMPORTANT than getting some successes. Because SUCCESSES ARE WHAT MATTERS. Something that is enshrined in NGRouting. Local node thread load is not really what matters. Killing potentially successful transfers is INSANE, and I will revert any change made in CVS to that effect. It is *so* difficult to get to the transferring data stage, killing them is monstrously counterproductive. And finally, because it would be exploitable. Even more than the current mechanism. An attacker can DoS a node with very little bandwidth usage by starting a bunch of trickled inserts to the node. This will be fixed by NIOv2. But with your scheme, he can even kill off the old requests - and then simply stop sending and let his transfers time out. He can then move on to another node. -- Matthew J Toseland - [EMAIL PROTECTED] Freenet Project Official Codemonkey - http://freenetproject.org/ ICTHUS - Nothing is impossible. Our Boss says so.
pgp00000.pgp
Description: PGP signature
_______________________________________________ Devl mailing list [EMAIL PROTECTED] http://dodo.freenetproject.org/cgi-bin/mailman/listinfo/devl
