Re: [PATCH 2/3] sched: only run newidle if previous task was CFS

[Peter Zijlstra]

On Tue, 2008-06-24 at 11:58 +0200, Peter Zijlstra wrote:
> On Mon, 2008-06-23 at 17:04 -0600, Gregory Haskins wrote:
> > A system that tends to overschedule (such as PREEMPT_RT) will naturally
> > tend to newidle balance often as well.  This may have quite a negative
> > impact on performance.  This patch attempts to address the overzealous
> > newidle balancing by only allowing it to occur if the previous task
> > was SCHED_OTHER.
> > 
> > Some may argue that if the system is going idle, it should try to
> > newidle balance to keep it doing useful work.  But the fact is that
> > spending too much time in the load-balancing code demonstrably hurts
> > performance as well.  Running oprofile on the system with various
> > workloads has shown that we can sometimes spend a majority of our
> > cpu-time running load_balance_newidle.  Additionally, disabling
> > newidle balancing can make said workloads increase in performance by
> > up to 200%.  Obviously disabling the feature outright is not sustainable,
> > but hopefully we can make it smarter. 
> > 
> > This code assumes that if there arent any CFS tasks present on the queue,
> > it was probably already balanced.
> > 
> > Signed-off-by: Gregory Haskins <ghaskins@xxxxxxxxxx>
> 
> NAK, this wrecks idle balance for any potential other classes.
> 
> idle_balance() is the generical hook - as can be seen from the class
> iteration in move_tasks().
> 
> I can imagine paritioned EDF wanting to make use of these hooks to
> balance the reservations.

Hmm, it wouldn't,.. since its too tied in with fbg which is sched_other
based,..

would need more generalization work,..



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/