I've been experiencing and documenting a pretty severe performance problem with SA versions 3.0.1 through 3.1x (nightly) under Solaris 8 and 9, Perl 5.8.3.
We're running Sendmail 8.12.11, and I've tried milters MimeDefang and Spamass-milter.
I initially thought this problem was related to the "round-robin" forking of spamd, but find that the 3.1 nightly exhibits the same behavior (using the new pre-forking algorithms), regardless of the number of spamd children (which MimeDefang doesn't appear to use anyway).
My problem is that when running a test load of about 350 messages through a test box (serially, so we're only talking 1 message at a time), I see the CPU load peg out frequently, around 90% of the time. At other times during this cycle, the CPU load is between 30 and 70 % which I'd call acceptable. In our production environment, in which processing is not serial (multiple sendmail threads running), the CPU load kills the machine dead in short order.
When MimeDefang is used as the milter, I see a fairly even spread of CPU usage between user, system, and wait times. With the spamass-milter, I see almost all of the CPU consumed by user processing.
The SA results look much "better" when we use spamass-milter / spamd, as I think MimeDefang doesn't round the scores up.
I'm wondering if there's some kind of Perl or Solaris "tuning" that I might need to do in order to not kill the CPU so bad. I've tried "nice"ing spamd, but that really didn't do much for the problem.
Anyone have any ideas or suggestions of places to look?
Thanks!
