On Wed, Dec 14, 2011 at 2:10 PM, Emmanuel Lecharny <elecha...@gmail.com> wrote: > Hi, > > we discussed off line with Julien about Idle session support. Here is a sum > up of what we talked about, and we may have to make some choice. > > First of all, it's absolutely mandatory to be able to manage idle sessions, > in order to kill them, for instance. So we must offer a mechanism by which > the application developer can handle this. > > In MINA 2, if a session is idle for more than a given amount of time > (configurable), either on read, write or both, then an event is fired on the > chain, up to the handler. > > This is done in the selector loop. The problem with this approach is that > it's extremely expensive, and most certainly the worst possible solution to > deal with idle sessions, when it comes to manage hundred or thousands of > idle sessions. Another issue is that because it's really awfully costly, the > quantum between two detection is one second. Last, not least, this one > second os not very precise for a session, as if the session was created > between two checks, then the delay between the last check and the session > creation adds up. In other words, we can wait up to one more second to > detect that a session is idle. > > We have some better ideas : > - first, not all the sessions can be controlled. We need a list of session > registered for an idle check > - second, and more important, we must have a dedicated thread > - last, not least, this mechanism should be optional. Ie, when the session > is created, it's up to the application to tell that the session must be > checked for idleness, by registering it into the idleController. > > We should also allow a more fine grained control, like 100ms, and even, it > could be made configurable. > > thoughts ? >
The current mechanism is pretty wrong, and cannot resolve idle time lesser than 1sec. We should do something not mandatory and efficient, I'm thinking about a time-period HashMap of list of session to check (we should update it for each session I/O events), and a timer checking every X ms the list of session in the map for this time index. Well I 'll try to clear my mind a propose a design. Julien
