On Thu, Aug 05, 2004 at 09:42:16PM -0700, Aaron J. Grier wrote:
> my 2x CPU sparc 20 has been up for 57 days now running
> 
> NetBSD radbug 2.0_BETA NetBSD 2.0_BETA (GENERIC.MP) #0: Thu Apr 29 14:07:17 UTC 2004 autobuild@tgm.netbsd.org:/autobuild/netbsd-2-0/sparc/OBJ/autobuild/netbsd-2-0/src/sys/arch/sparc/compile/GENERIC.MP sparc
> 
> with spamassassin-2.63nb1 and perl-5.8.4.
> 
> the last few days spamassassin has been flailing wildly giving me out of
> memory messages:
> 
> sbrk: grow 372736 failed, error = 12
> 
> and leaving me with piles of unfiltered spam.
> 
> it doesn't seem to be a ulimit issue; I have an i386 with even less
> memory, configured with a week-older kernel version, the same versions
> of spamassassin and perl.   given the same data files, and the same
> ulimits, it doesn't run out of memory.
> 
> the issue appears to be sparc-specific.  since this has only cropped up
> over the last couple days, as spamassassin was running fine previously,
> I'm wondering if there's not some memory fragmentation issue.
> 
> compiling up a new kernel I get
> 
> --- ip_nat.o ---
> virtual memory exhausted: Cannot allocate memory
> 
> does this look familiar to anybody?

Yes, I believe so -- I see "virtual memory exhausted"... messages
fairly regularly on both my SS20.  As you say, mostly during builds.

I'm not running spamassassin on either, fwiw -- the 2.0_BETA GENERIC 
ss20 is running bind and sendmail and that's about it.  The 2.0_BETA
ss20 w/GENERIC.MP is the "hot spare" and spends its time building
kernels and world.

I'm not sure how to help isolate either (suggestions welcome, of
course), but I think I would agree w/your notion about suspecting
memory leakage/fragmentation -- both ss20 have 512MB and were
performing their respective duties running 1.6 for months without
incident.  I was excited (still am, really) about running MP in
these systems, hence the upgrade to 2.0 .

I'm doing new CVS/kernel/world today on the "hot spare" just
to start fresh again.

sr.