vmstat's w column

[Michael W. Lucas]

On Wed, Oct 04, 2017 at 09:12:26AM +0300, Konstantin Belousov wrote:
> On Tue, Oct 03, 2017 at 02:57:28PM -0700, Don Lewis wrote:
> > On  3 Oct, Michael W. Lucas wrote:
> > > Hi,
> > > 
> > > (This query brought to you courtesy of tech reviews on my new freeBSD
> > > book.)
> > > 
> > > On my -current box, vmstat(8) says:
> > > 
> > >      procs   Information about the numbers of processes in various states.
> > > 
> > >              r       in run queue
> > >              b       blocked for resources (i/o, paging, etc.)
> > >              w       runnable or short sleeper (< 20 secs) but swapped
> > > 
> > > I've had a couple people report to me that they have w entries even
> > > when only a few kilobytes are swapped out.
> > > 
> > > Is the man page wrong? Or should I tell them to report that output as
> > > a bug?
> > 
> > I don't know about the runnable or short sleeper bit.  On currently idle
> > package build box vmstat currently reports 15 processes in the w column.
> > If I look at the output of ps, there are 15 processes with a state of
> > either IW or TW, all of which have a RSS of zero.  These have probably
> > all been idle since before the end of the last poudriere run, which is a
> > very heavy swap user.  There are no processes in an RW or SW state.  It
> > looks like vmstat is reporting all swapped processes, so the man page
> > and implementation appear to be out of sync, at least on recent
> > 12.0-CURRENT.  My other two currently running FreeBSD boxes don't have
> > any swapped processes reported by either ps or vmstat.
> 
> Present definition of 'swapped out' is a process for which all threads
> have their kernel stacks swapped out.  Swapper does not perform any
> efforts to swap out userspace memory of the process.  The reasoning is
> perhaps that normal pagedaemon processing would inactivate/page out
> the unreferenced pages on busy system, and swapping the kernel stacks
> never occurs if pagedaemon can keep with the memory demand without such
> drastic measures.
> 
> In other words, RSS == 0 is neither necessary nor sufficient to conclude
> that a process is swapped out.

Thanks, folks. I'll be adding bits of this conversation to the
swapping discussion.

And I'm hearing, from an end users' point of view:

The w column no longer provides useful performance information.

==ml

-- 
Michael W. Lucas 	https://mwl.io/
nonfiction: https://www.michaelwlucas.com/
fiction: https://www.michaelwarrenlucas.com/