Re: 2.6.8-rc2-mm2

[Con Kolivas]

Helge Hafting wrote:
> Con Kolivas wrote:
>
> > Ingo Molnar writes:
> >
> > > * Con Kolivas <kernel@xxxxxxxxxxx> wrote:
> > >
> > > > Can you define them please? I haven't had any reported to me.
> > >
> > >
> > >
> > > sure: take a process that uses 85% of CPU time (and sleeps 15% of the
> > > time) if running on an idle system. Start just two of these hogs at
> > > normal priority. 2.6.8-rc2-mm2 becomes almost instantly unusable even
> > > over a text console: a single 'top' refresh takes ages, 'ls' displays
> > > one line per second or so. Start more of these and the system
> > > effectively locks up.
> >
> >
> >
> > It's only if I physically try and create such a test application that 
> > I can reproduce it. I haven't found any real world load that does that 
> > - but of course that doesn't mean I should discount it. However, 
> > interactive mode off doesn't exhibit it and it should be easy enough 
> > to fix for interactive mode on. Thanks for pointing it out.
>
>
> I can easily imagine this sort of real-life application.  Anything that 
> generate real-time data, such as sound or video for immediate playback 
> will use some continous fraction of cpu, depending on how much data and 
> how fast the cpu happens to be.   A scheduler with such
> a problem will now and then cause trouble when some processor needs just 
> the
> "wrong" percentage of its capacity to run the task(s). Consider someone 
> playing several video files until he exceeds the capacity of the 
> machine. Video skipping frames is then ok, a frozen UI is not.  Or 
> consider a multiuser machine running several games.  Or even a parallell 
> kernel compile when the source isn't cached. The compiler(s) will work 
> for a while, wait for disk for a while, work for a while . . .

Your concerns are valid, and are relevant for _any_ scheduler design. 
However none of these occur with staircase. Ingo is describing a special 
case where it runs for more than 1ms and less than 10ms. Most things do 
not run exactly in this manner, which is why you haven't seen it, and 
neither have I till I created his test case.

As for the multiuser machine, this is actually an area where staircase 
_excels_ at when put into non-interactive mode. The fact that 
interactivity is not as good in interactive mode under load in a single 
desktop mode doesn't really describe that interactivity is still 
remarkably good considering how strict the cpu distribution is - and 
that's by design. I have now a couple of users using staircase in 
non-interactive mode sharing resources between up to 10 users happily, 
where previously the mainline 2.6 kernel was unable to do this in a fair 
manner.

> 2.6.8-rc2-mm2 seems to work fine for me right now, but I hope there will 
> be solutions for this sort of problem.

Great.

> Helge Hafting

Cheers,
Con

Attachment: signature.asc
Description: OpenPGP digital signature