[exim] Running exim on a cluster-filesystem

Top Page
Delete this message
Reply to this message
Author: Patrick von der Hagen
Date:  
To: exim-users
Subject: [exim] Running exim on a cluster-filesystem
Hi all,

I'm trying to run a cluster-setup where a shared filesystem is providing
a central location for configuration and a common spool to several exims.

Basically, a big filesystem is accessed concurrently by three
mail-servers which would probably be impossible if exim used IPC,
shared-memory or other "local" means of communication between the single
exim-processes. However, since exim just does not do that and since the
filesystem claims to have working locking-mechanisms, this sounds like a
possible setup.

However, some files, like the hints-databases, might prove to cause
performace-problems.

So far, I just run exim as queue-only without any queue-runners, in
order to gather the best-case results. Using three P3-550 with 256MB RAM
each (certainly no fast setup) I get about 14 mails/second if they use
the same filesystem but different spool-directories (/var/test1/spool,
/var/test2/spool, /var/test3/spool). I don't really care wheter this is
particularly good or bad, given my specific hardware. Running the same
test on a local harddisk gives me 15 mails/second, so the remote
filesystem doesn't perform too badly.

Now the problem: using the same spool-directory on all servers and no
other changes, the throughput drops to less than 6 mails/second on my
servers, giving me a total of about 17 mails/second. One server makes
14, three servers get 17.... no scalability at all.

First idea: some locking issue. But what can I do, where can I start
profiling?
Second idea: might enablich splitspool help?

Well, splitspool does help. It gets 8,5 mails/second and thus performs
much better than a normal spool. However, overall performace still is
not what I had hoped for....

So, please share your advice with poor littel me. ;-)
Would you suggest that the complete setup is going to go BOOOM and I
have missed some evil pitfalls?
Would you think that the performace-problem is exim-related? Especially
file-locking-related?
Or would you suggest that my filesystem is causing me headaches, since
the big impact splitspool has might indicate problems if there are too
many files in a directory? When I perform my tests I usually send
several thousand mails, so there tends to be a lot of files in the spool....

-- 
CU,
    Patrick.