Re: [Exim] No locking problem for hints databases

On Wed, Oct 15, 2003 at 10:46:46PM +0200, Michael Haardt wrote:
> That's quite different from your setup, but since I raised the question if
> avoiding the hints might speed up local delivery, I thought I should also
> try to answer that, as it might save Philip some work that has no effect.

Yes, local delivery wrt the *hints* database has never really been a
problem.
The problem I and I think some others were seeing is with the
wait_smtp-remote database, and being limited to ~10,000 remote
deliveries per hour on fallback hosts (hosts queuing messages which
couldn't be delivered first time).

> If you can not deliver most messages on first try, how about not trying
> direct deliveries and leaving that to queue runners?

Because on high throughput relays, having growing queues is inefficient
and slows the system down :)
It's mor efficient to pass messages which cannot be delivered first
time, to a seperate host to queue.

> > At the same time, were there new messages coming in to the queue?
>
> Yes, but not many. Check src/dbfn.c and grep for F_SETLKW to find where
> the DBs are locked. It is easy to add some additional code there. I
> also suggest to use tmpfs for the DBs. It takes 30 MB on my outgoing
> servers, and much less everywhere else.

I'm currently playing with removing the wait_smtp-remote databases all
together and seeing what that does to throughput.
It might help, but there would be the obvious disadvantage that not all
messages to a single host would get pipelined away first time when the
host returns.