Re: [RFC 4/7] change kernel accounting to include steal time

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Aug 26, 2010 at 05:28:56PM -0300, Glauber Costa wrote:
> On Thu, Aug 26, 2010 at 02:23:03PM -0300, Marcelo Tosatti wrote:
> > On Wed, Aug 25, 2010 at 05:43:14PM -0400, Glauber Costa wrote:
> > > This patch proposes a common steal time implementation. When no
> > > steal time is accounted, we just add a branch to the current
> > > accounting code, that shouldn't add much overhead.
> > > 
> > > When we do want to register steal time, we proceed as following:
> > > - if we would account user or system time in this tick, and there is
> > >   out-of-cpu time registered, we skip it altogether, and account steal
> > >   time only.
> > > - if we would account user or system time in this tick, and we got the
> > >   cpu for the whole slice, we proceed normaly.
> > > - if we are idle in this tick, we flush out-of-cpu time to give it the
> > >   chance to update whatever last-measure internal variable it may have.
> > 
> > Problem of using sched notifiers is that you don't differentiate whether
> > the vcpu scheduled out by its own (via hlt emulation) or not.
> And we don't need to. If we're out because we want to, we're idle.
> And so, we don't account steal time.

Think of the program below.

> > Skipping accounting of user/system time whenever there's any stolen
> > time detected probably breaks u/s accounting on non-cpu-hog loads.
> I am willing to test some workloads you can suggest, but right now,
> (yeah, I mostly used cpu-hogs), this scheme worked better.
> 
> Linux does statistical sampling for accounting anyway, so I don't see
> it getting much worse.

A "cpu hog" that sleeps 1us every 1ms.

> > I suppose steal time should be accounted separately from u/s ticks, as
> > Xen does.
> It requires us to hook somewhere else, which I deem as overcomplicated.
> Do you have any suggestion on how to make it simple?

Unfortunately no.

> Furthermore, "doing separate", is equivalent of not skipping user/system,
> if we really prefer to.
>
> > +   if (delta > 1000UL)
> > +               touch_softlockup_watchdog();
> > +
> > 
> > This will break authentic soft lockup detection whenever qemu processing
> > takes more than 1s.
> 
> This should be 10s. 1000UL is a typo.

Comment is still valid.