Re: Scheduler(?) regression from 2.6.22 to 2.6.24 for short-lived threads

[Olof Johansson <olof@lixom.net>]

On Mon, Feb 11, 2008 at 08:58:46PM +0100, Mike Galbraith wrote:
> 
> On Mon, 2008-02-11 at 11:26 -0600, Olof Johansson wrote: 
> > On Mon, Feb 11, 2008 at 09:15:55AM +0100, Mike Galbraith wrote:
> > > Piddling around with your testcase, it still looks to me like things
> > > improved considerably in latest greatest git.  Hopefully that means
> > > happiness is in the pipe for the real workload... synthetic load is
> > > definitely happier here as burst is shortened.
> > 
> > The real workload doesn't see much of an improvement. The changes I did
> > when tinkering yesterday seem like they're better at modelling just
> > what's going on with that one.
> 
> So the real application is trying to yield?  If so, you could try
> prodding /proc/sys/kernel/sched_compat_yield.

The real application wasn't using yield, but it also didn't get hung up
on the busy-wait loop, that was a mistake by me on Friday. Practically
all the time of that application was spent in the actual workload
loop. I tried adding a yield at the bottom loop, but it didn't make a
significant difference.

> It shouldn't matter if you yield or not really, that should reduce the
> number of non-work spin cycles wasted awaiting preemption as threads
> execute in series (the problem), and should improve your performance
> numbers, but not beyond single threaded.
> 
> If I plugged a yield into the busy wait, I would expect to see a large
> behavioral difference due to yield implementation changes, but that
> would only be a symptom in this case, no?  Yield should be a noop.

Exactly. It made a big impact on the first testcase from Friday, where
the spin-off thread spent the bulk of the time in the busy-wait loop,
with a very small initial workload loop. Thus the yield passed the cpu
over to the other thread who got a chance to run the small workload,
followed by a quick finish by both of them. The better model spends the
bulk of the time in the first workload loop, so yielding doesn't gain
at all the same amount.

I still added it to rule out that it was a factor in the time
differences.


-Olof