did anybody ever try multiple exim processes that accept() connections
from the same socket? It would reduce the latency on heavily loaded
systems, but I have no clue if that would make any difference for mail
systems.
While talking about heavy load: As the recent discussion showed, Exim
is not made for huge queues. Although it is a hack, I suggest the
following feature (for discussion, so far):
There should be a delay time, after which a further queue runner will
be started, if the last started queue runner has not yet exited during
that time interval.
During times of heavy load, where lots of unprocessed mail sits in the
queue, this feature should make sure that the needed number of queue
runners is reached earlier. Somehow I think that mails that were just
queued and not deferred should be given precedence, but so far I can't
decide how. First scan the queue, then deliver fresh mails, then deliver
deferred mails? That would introduce unneeded queue scanning overhead.
Have separate queue runners for fresh mails? That would make it even
harder to balance the system. Reorganise the queue structure, which
effectively introduces a priority scheme? *hides and runs* :)
