When I look at 'top', the spamd processes are not sharing enough memory

Here's a sample output from "top" viewing two spamd processes:

  PID USER      PR  NI CODE  VIRT  RES  SHR nDRT %CPU %MEM   TIME COMMAND
 7158 jm        16   0  964 34300  31m 2280    0  1.8  3.1   0:02 perl
 7409 jm        17   0  964 34300  31m 2280    0  0.0  3.1   0:00 perl

Looking at the above 'top' output, it appears that the two spamd processes listed are sharing a pitiful 2280KB of memory pages – 3.4% of their total memory usage. However, all's not as it seems.

...

What's happening is that, sometime during kernel 2.5 development, the memory management subsystem was changed to no longer track data pages shared between multiple processes through copy-on-write, as 'shared'. Instead, only text pages, and pages loaded from shared libraries, are considered part of the 'shared' section, in contradiction to what the top(1) manual page states.

This is generally observable in all 2.6.x kernels and Red Hat's patched versions of the 2.4.1x and 2.4.2x kernels. Reportedly, RHEL 2.1 ES and AS kernels are the same, too.

...

In reality, each spamd process is probably sharing somewhere between 50% and 90% of its pages with the other spamds (depending on its age), but 'top' is unable to report this correctly.

Wiki MarkupMore details, and a demonstration script, \[http://bugzilla.spamassassin.org/show_bug.cgi?id=3839#c10 can be found here\].

Workarounds

If you want a reliable idea of your memory sharing, so far the only one way to determine this seems to be to downgrade the kernel to vanilla 2.4.\[12\]\* or test on a non-Linux platform.  There's also patches that can be applied to 2.6 to get useful figures via the /proc/$$/smaps interface; [lkml thread on the topic|http://lkml.org/lkml/2005/7/6/250].