Can you reproduce it easily with a simple test program that connects/disconnects to SSL port quickly?
On Fri, 26 Jan 2007, Matt Mankins wrote: > > 3) The process was using 100% of CPU and taking up 126MB. The > > number of active connections was less than 10. It sits behind LVS > > which checks for connection uptime periodically which is why the > > high connection count. > > Incidentally, an strace on the djabberd process shows a zillion: > > epoll_wait(9, {{EPOLLIN, {u32=13, u64=13}}}, 1000, -1) = 1 > epoll_wait(9, {{EPOLLIN, {u32=13, u64=13}}}, 1000, -1) = 1 > epoll_wait(9, {{EPOLLIN, {u32=13, u64=13}}}, 1000, -1) = 1 > > I wonder if the watch_read or watch_write is getting messed up when > LVS does a connect and then disconnect on the SSL port? It seems to > build over time and the watch on/off with SSL looks tricky. > > > > > > >