The detailed analysis can be found at:
The thing is that IoServiceListenerSupport.fireSessionCreated() and fireSessionDestroyed() can be invoked at the same time by more than one I/O processor threads, and that can cause sessionClosed event is not fired at all under heavy load. The solution is to acquire a per-service lock whenever managed session set changes. So far, we were acquiring locks separated as two phases, and that was the root cause of the problem. By fixing this issue, performance might decrease slightly, but accuracy is the first concern for us.