Re: 2.6.19-rc5: known regressions

[Tim Chen <tim.c.chen@linux.intel.com>]

On Wed, 2006-11-08 at 15:11 -0800, Tim Chen wrote:
> On Wed, 2006-11-08 at 17:22 +0100, Adrian Bunk wrote:
> 
> > There's perhaps one thing that might help us to see whether it's just a 
> > benchmark effekt or a real problem:
> > 
> > With Tim's CONFIG_NR_CPUS=8, NR_IRQS only increases from 224 in 2.6.18 
> > to 512 in 2.6.19-rc.
> > 
> > With CONFIG_NR_CPUS=255, NR_IRQS increases from 224 in 2.6.18
> > to 8416 in 2.6.19-rc.
> > 
> > @Tim:
> > Can you try CONFIG_NR_CPUS=255 with both 2.6.18 and 2.6.19-rc5?
> > 
> 
> With CONFIG_NR_CPUS increased from 8 to 64:
> 2.6.18     see no change in fork time measured.
> 2.6.19-rc5 see a 138% increase in fork time.
> 

Lmbench is broken in its fork time measurement.
It includes overhead time when it is pinning processes onto
specific cpu. The actual fork time is not affected by NR_IRQS.

Lmbench calls the following C library function to determine the 
number of processors online before it pin the processes: 
	sysconf(_SC_NPROCESSORS_ONLN);

This function takes the same order of time to run as
fork itself.  In addition, runtime of this function 
increases with NR_IRQS.  This resulted in the change in
time measured.

After hardcoding the number of online processors in lmbench,
the fork time measured now does not change with CONFIG_NR_CPUS
for both 2.6.18 and 2.6.19-rc5.  So we can now conclude that
NR_IRQS does not affect fork.  We can remove this particular
issue from the known regression.

Tim