kernel builds starving evolution process - scheduler issue? (was Re: [patch] Real-Time Preemption, -RT-2.6.10-rc2-mm2-V0.7.30-9)

[Lee Revell <rlrevell@joe-job.com>]

On Wed, 2004-11-24 at 08:33 -0500, Steven Rostedt wrote:
> On Wed, 2004-11-24 at 04:45 +0100, Ingo Molnar wrote:
> > * Lee Revell <rlrevell@joe-job.com> wrote:
> 
> > > 
> > > The symptom is that CPU bound tasks like kernel compiles will starve
> > > I/O bound tasks like evolution for a _long_ time.  If I have a kernel
> > > build and external modules building at the same time and Evolution
> > > goes to "Update message list...", it can sit and spin with a blank
> > > message pane for a minute or two.  If I suspend the builds, the
> > > message list renders immediately.
> > 
> > could you try the vanilla -rc2-mm2 kernel (with PREEMPT enabled), does
> > it behave in such a way too? At first sight this could be a property of
> > the upstream scheduler, but maybe it's special to PREEMPT_RT.
> > 
> 
> Have you notice this behavior with other interactive (I/O) tasks, such
> as bash.  Evolution is quite a big utility, and might be doing something
> in the background. If you see the same behavior with bash then there is
> no doubt that the compile is slowing down an I/O intensive task.
> 

No.  Only evolution (2.0) exhibits the problem.  But, it looks like
evolution uses a comparable amount of CPU to a kernel build just
updating the message list.  All stracing it shows me is that it spends a
hell of a lot of time polling().  I think this might be a bloat issue.

> Another variable can be memory. Are you running this on something with
> adequate memory, or is you harddrive churning like mad and you're
> constantly thrashing the swap space?
>  

No, I have plenty of RAM (512M).  I am using a 600Mhz C3 so the system
is probably CPU bound.  But, it seems like evolution should make a
little more progress.  I often find myself having to background all
build processes for a few seconds to let the message list render.  Once
the list renders, and I resume the builds, evolution is more or less
usable.  Running gtk-gnutella in the background will also make evolution
horribly slow.

Running the offending, CPU bound processes at a high nice value solves
the problem.  But now I am wasting half my fscking cycles "Updating
message list...".   Grr.

Lee