Invalid memory stats from vmstat and sysctl vm.vmtotal?

[Jason Hellenthal]

On Fri, Dec 02, 2011 at 05:13:05PM -0000, Steven Hartland wrote:
> ----- Original Message ----- 
> From: "Andriy Gapon" <avg at FreeBSD.org>
> 
> >> Totalling up RSS from ps axo "rss" gives a total in the region of that if
> >> the vm stats are out by a factor of 4, in this case it should be: 8132557
> >> which is 7.75GB a much more realistic value.
> >> 
> >> Am I totally missing something or is there problem here?
> > 
> > Likely more of the former than of latter.  Those virtual sizes are not
> > sufficiently explained, but you have been warned that those are not physical
> > sizes, so I am not sure why you try to compare the virtual figures with the
> > physical figures.
> 
> My miss-understanding was due to what "virtual" actually meant.
> 
> > Here's an example.  Let' say you mmap-ed a 1GB file into a process memory space,
> > here you immediately increased your virtual size counts by 1GB, even if you hadn't
> > accessed any bytes in the file yet and so none of them were in physical memory.
> > The same applies to anonymous memory.
> > 
> > P.S. the above is reveled by a cursory look through the code (which is publicly
> > available btw) :-)
> 
> Yer I did have a dig around before posting and ended up the code for vm.vmtotal,
> which is where vmstat gets its info from but that's just a summation of each object's
> size from vm_object_list. Thats where I got lost without an insight into what
> a vm_object.size actually represents.
> 
> Your info about mmap'ed files helped point me in the right direction as it 
> identified space that shows as virtual but doesn't show in swap or real ram,
> which is what I was missing.
> 
> Given this starting point the following links provided me with addtional
> information:-
> http://www.freebsd.org/doc/en/books/arch-handbook/vm.html
> http://www.freebsd.org/doc/en/books/design-44bsd/overview-memory-management.html
> http://www.freebsd.org/doc/en_US.ISO8859-1/articles/vm-design/
> http://www.cse.chalmers.se/edu/course/EDA203/unix4.pdf
> 
> I was under the incorrect impression that Virtual Memory (VM) was so named as it
> was a unified physical memory and swap (virtual memory), but its not that simple,
> as other items such as file-backed objects also count to this total which would
> never show in physical or swap allocation of other tools such as top and swapinfo.
> 
> So what I believe is now the big cause of virtual memory uplift vs the memory
> totals shown by ps / top is that the vm totals include things like file backed
> memory mapped process binaries, shared libs etc many multiple times.
> 
> This would explain why this specific machine shows the applification more than
> others here as it runs thousands of very small lightweight processes.
> 
> Thanks for pointer Andy, I now understand a lot more about the BSD VMS :)
> 
> What do people think about expanding that entry in the man page of vmstat to
> clarify just what active "virtual pages" really means?
> 
>     Regards
>     Steve
> 

Thanks for your research Steve. That makes perfect sense and additions to the documentation are surely needed.