Tavin: I wouldn't mess with the thread code until the after we do a stable 0.5 release. Without fail, almost every single "simple" "obvious" improvement to the threading system or OCM has resulted in subtle bugs. Adding more new untested code to a codebase never increases stability.
My priorities as we move towards 0.5 would be: 0) Stabilize the new data store implementation. It seems too hard to request large (>32K) chunks of data from the network. It smells like there are still ring buffer bugs that are impacting tunneling. In the history of 0.4 Fred, we have never had a stable and robust Datastore implementation. You have done a lot of good work. Please don't get distracted before the job is finished. 1) Address simple parameter tuning. If it's true that the hop times are pathologically misconfigured, we might be able to get a large improvement in network performance just by fixing them. Likewise, the thread pool sized based QueryRejection mechanism hasn't been retuned since inbound connection threads started being pulled out of the thread pool again. We probably need to make the default thread pool size bigger. 2) Characterize load balancing issues. Apply quick fix if there is one. I have yet to get an answer from anyone on what node throughput needs to be. I assume that no one knows the answer to this question. If there's positive feedback in the network that causes nodes to be driven to overload, increasing the the throughput rate won't help -- nodes will just saturate at the new higher rate. My experience is that my nodes seem to get overloaded (~200 - 400 requests /minute, ~= 100 inbound connections a minute) after about 24 hours of operation. I think Oskar was talking about doing full simulations of different load balancing strategies. I don't know how long that will take. In the meantime we should at least keep track of the node's load and make sure that it stops giving out references to itself when it's overloaded. --gj -- Freesite (0.4) freenet:SSK at npfV5XQijFkF6sXZvuO0o~kG4wEPAgM/homepage// _______________________________________________ Devl mailing list Devl at freenetproject.org http://lists.freenetproject.org/mailman/listinfo/devl
