Re: [Lse-tech] Re: [PATCH] [PERFORMANCE RESULTS] priority preemption in Linux

[Bill Hartner <hartner@austin.ibm.com>]

Ingo Molnar wrote:
> 
> On Tue, 15 Oct 2002, Bill Hartner wrote:
> 
> > kernel    preemption     conforming     percent
> >           threshold      connections    improvement
> > ======    ===========    ===========    ===========
> >
> > 2.5.33    0 (default)    2906           baseline
> > 2.5.33    40             2990           2.9 %
> >
> > Table 3 shows that reducing priority preemption improved the
> > number of conforming connections by 2.9 %.
> 
> actually, the more interesting metric is the ops/sec value - how did that
> change? Conforming connections is a cutoff value and the real improvement
> might be bigger than that.

The 2.5.33 baseline was 7952 ops/s and it improved to 8101 which is 1.9 %.
The SPECweb99 UP needs to be instrumented to see what the level of priority
preemption is.  May not be a problem at all - the VolanoMark case had a group
of processes running at a higher priority and long run queues.

Bill