On Tue, 30 Apr 2002, Sheldon Hearn wrote:
> Well, it certainly does what I asked for. :-)
Glad to be of service.
> Given this, it's probably just going to be simplest to keep a separate
> queue for Hotmail (and any other problematic domains that crop up) and
> use fewer queue runners to process that queue.
Indeed.
> It still seems weird to me that as few as 600 queue runners can run into
> lock contention on a RAM-backed DBM file as small as 3MB.
Yes, it may be worth trying to find out what's going on here, in case
there is something that can be done in Exim. I thought I had written it
to minimize the time it holds a write lock, but ...
> I was looking at transport_update_waiting() and wondered whether I
> couldn't limit the write lock window, by downgrading the initial
> dbfn_opn() to O_READONLY and only acquiring a O_RDWR lock as and when
> required in the host_record-based for loop.
Possibly.
> I had a little trouble profiling this, though, and wondered if you could
> give me an "author's impression" of whether this is likely to reduce
> average write lock wait time at all. :-)
I'm "away" at the moment - no time to consider the code just now, sorry.
I know there are complications about changing from read to write locks,
and would have to check up on that.
You may just have to experiment by yourself. :-)
--
Philip Hazel University of Cambridge Computing Service,
ph10@??? Cambridge, England. Phone: +44 1223 334714.
