On Fri, Mar 08, 2013 at 07:09:12PM +0200, Henrik K wrote: > On Fri, Mar 08, 2013 at 11:55:43AM -0500, David F. Skoll wrote: > > On Fri, 8 Mar 2013 18:44:54 +0200 > > Henrik K <h...@hege.li> wrote: > > > > > Virgin childs: > > 79MB private; 2GB shared (~40MB shared/child) > > > > > Used childs: > > 1.2GB private; 1.2GB shared (~24MB shared/child) > > > > This is roughly what I am seeing with MIMEDefang also: Only 50% > > shared. It's better than nothing, but not nearly as good as one might > > have hoped. On my system, about 3.5MB of the shared memory is the > > text portion of libraries such as libperl.so and libc.so which you'd > > expect to be shared anyway, so really only about 20.5MB/46.5MB of Perl > > memory is shared for each slave. > > But atleast ~10-20MB of private data per spamd child is the per message scan > data/blobs etc (can be seen as 20mb heap allocation). This should not be > calculated in any memory ratio conclusions.
Also to be noted that these "shared" values are completely vague anyway. All that matters is how much real system memory is taken. As seen by my previous free reports, master + 50 virgin childs only take total of 137MB. At full blast everything take ~1.2GB. So the real ratio per child might be something like 4MB bogus perl data, 20MB of per-message data. If the master process takes 50MB memory, this means the child ratio is _much_ better than 50% (4MB/50MB). I might dig deeper into this later with some tools.
