Re: Locking and concurrency -- system considerations

Kezdőlap
Üzenet törlése
Válasz az üzenetre
Szerző: Philip Hazel
Dátum:  
Címzett: Jon Morby
CC: exim-users
Tárgy: Re: Locking and concurrency -- system considerations
On Wed, 24 Sep 1997, Jon Morby wrote:

> Individual logs isn't such a problem, although it would be nice to have a
> central logging daemon which collated all the logs (a la syslog, only more
> reliable) so you could have one large central log file.


I'm afraid I wrote Exim to be simple. While being aware of the fact that
one's programs always get used for tasks far bigger than you ever
thought of (I've learned this the hard way over the last 30 years), I'm
still amazed at the sort of things people fling at Exim (messages with
1000's of recipients and the like). I never envisaged it as an engine
for use in extreme environments. I guess I suffer from imagination
failure. :-)

> In the environment I envisage running Exim shared spools are essential. The
> current system we have uniques the spool file by taking the numeric ID from
> the hostname and using that as well as the pid and current time.
>
> we have several machines (relay-1 -> relay-15 currently) all of which share
> one of two spools.
>
> Being restricted to one spool per delivery client would be a restriction I
> wouldn't like to have to live with.


The fact that message ids are 16 characters long is probably assumed in
too many places in Exim to make it easy to change. The current id is
made up of 6 characters of time, in seconds, converted to base 62,
another 6 characters of pid, also to base 62, and two digits which are
nearly always 00, but can vary if a fast machine receives more than one
message per second in the same process:

tttttt-pppppp-nn

Now, the 6-digit base-62 pid number can actually hold very large
numbers. A 3-digit base-62 number can hold up to 238327, which is
sufficient to hold a pid in most machines. One could therefore envisage
using the top 3 digits to hold some kind of machine identification. Even
if 4 digits were used for the pid (allowing up to 14776335 for the pid)
that would still leave two digits for the machine id - 3843 possible
values. Modifying Exim to generate message ids in this way would not be
a big task.

Another place where a modification would be required is when
it is writing the header spool file. This is written with a temporary
name and then renamed. Incorporating something machine-specific into the
temporary name would not be too hard.

> It would also be very useful to be able to share the retry/wait db's (probably
> essential).


As these files are already locked when in use, you might think they
could already be shared. However, the only locking is by fcntl() so I
suspect things would go wrong. Code would have to be added to cause Exim
to get a lock on a marker file *before opening a database*. However, I'm
less convinced on the essentialness of sharing these dbs, since they are
only hints.

> > Essentially, the assumption is that each host has its own spool and log
> > files.
>
> Unfortunately this assumption doesn't scale particularly well :(


Well, at least it's linear. :-)

> (What happens when one of your relays goes down, and you've got a couple of
> thousand messages in it's queue?) The mail gets delayed until you can bring
> another box up and mount that spool. Unacceptable in todays market with
> consumers rather than techies coming onto the Net.


How can I put this so that it doesn't sound like a whinge? You'll just
have to believe me when I say it isn't. I'm reacting to your comment:
If you are going to start talking market and consumers and commercial
presssures, then I am going to respond by making remarks about paying
commercial prices for your software. If it turns out to be possible to
convert Exim into a "massive, time-critical, serious commercial
environment" mailer, I'll be delighted, but that wasn't one of my
original design goals.

In our academic environment, if one of our relays goes down, then indeed
the mail does get delayed until we can bring up another box (if we
happen to have one). But then we live an an ivory tower...

Philip

-- 
Philip Hazel                   University Computing Service,
ph10@???             New Museums Site, Cambridge CB2 3QG,
P.Hazel@???          England.  Phone: +44 1223 334714



--
* This is sent by the exim-users mailing list.  To unsubscribe send a
    mail with subject "unsubscribe" to exim-users-request@???
* Exim information can be found at http://www.exim.org/