At 10:34 AM 5/23/01 -0400, you wrote: > > After seeing cpu-overloaded leaf servers becoming an ever-increasing > > problem I've spent some brain cycles thinking of a possible solution, and > > this is what I came up with. Flame and/or praise, have at it. > >Interesting solution...but it doesn't actually help you reduce the number >of open fds. Most users are connected for long-haul periods; simply >reducing the max connections for a particular class without actually >disconnecting some users won't free up more than, say, 5 fds. Also, >there is currently no statistic kept on how long the last PING took to >be returned, but that is a problem I have a possible solution for, and >just haven't had time to code up (or verify interoperability). This is not true, when a server is carrying thousands of clients , the turnover is *high*. I have tried the concept I described already in a live situation manually, by editing the Y: lines and rehashing, and it does work. From the top of my head it took between 5 and 10 minutes to lose 500 clients, on a server loaded with 5000 clients. /Arjen PS: somebody needs to set a proper Reply-To: header for this mailing list.
