On Fri, 21 Dec 2001, Mike Kravetz wrote:
> On Thu, Dec 20, 2001 at 02:57:55PM -0800, Davide Libenzi wrote:
> > On 21 Dec 2001, Momchil Velikov wrote:
> > >
> > > I'd like to second that, IMHO the RT task scheduling should trade
> > > throughput for latency, and if someone wants priority inversion, let
> > > him explicitly request it.
> >
> > No a great performance loss anyway. It's zero performance loss if the CPU
> > that has ran the woke up RT task for the last time is not running another
> > RT task ( very probable ). If the last CPU of the woke up task is running
> > another RT task a CPU discovery loop ( like the current scheduler ) must
> > be triggered. Not a great deal anyway.
>
> Some time back, I asked if anyone had any RT benchmarks and got
> little response. Performance (latency) degradation for RT tasks
> while implementing new schedulers was my concern. Does anyone
> have ideas about how we should measure/benchmark this? My
> 'solution' at the time was to take a scheduler heavy benchmark
> like reflex, and simply make all the tasks RT. This wasn't very
> 'real world', but at least it did allow me to compare scheduler
> overhead in the RT paths of various scheduler implementations.
Mike, a better real world test would be to have a variable system runqueue
load with the wakeup of an rt task and measuring the latency of the rt
task under various loads.
This can be easily implemented with cpuhog ( that load the runqueue ) plus
the LatSched ( scheduler latency sampler ) that will measure the exact
latency in CPU cycles.
- Davide
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sun Dec 23 2001 - 21:00:25 EST