Invert sense of SLB class bit

Sonny Rao sonny at burdell.org
Fri Sep 9 04:32:04 EST 2005


On Thu, Sep 08, 2005 at 02:57:11AM -0400, Sonny Rao wrote:
> On Thu, Sep 08, 2005 at 04:33:57PM +1000, Paul Mackerras wrote:
> > Sonny Rao writes:
> > 
> > > So, when exactly do we go into real mode ?  I know the SLB handler is
> > > now run in real mode, but is anything else? 
> > 
> > On every interrupt and exception (including system calls); i.e. every
> > time we enter the kernel.  A few are handled entirely in real mode,
> > but for most we do a small amount of exception entry code and then do
> > an rfid instruction to turn the MMU back on.
> > 
> > > If we want to look at best-case improvements -- meaning a lot of SLB
> > > reloading and userspace context-switching, I suppose I should select a
> > > machine with a fair amount of memory.
> > > 
> > > Either way, I'm going to run SDET on a 2-way box w/ 8GB of RAM and see
> > > if I can find a difference.  If necessary, I'll find a bigger box to
> > > test on.
> > 
> > SDET sounds like a good choice, since it will do lots of system calls
> > as well as context switches.

Ok here are the results for 1,4,16, and 64 procs on stock 2.6.13 and
2.6.13 plus the patch.

the tests were run three times and averaged with the std. dev in parens

Kernel	1 proc		4 procs		16 procs	64 procs
--	--		--		--		--
stock	4580 (29)	14757 (215)	15246 (244)	15317 (52)
patched	4556 (51)	14694 (116)	15320 (293)	15390 (38)

delta	-0.05 %		-0.04 %		+0.05 %		+0.05 %

In all cases except 64 procs, the std. deviation is bigger than the
difference.  

Sonny




More information about the Linuxppc64-dev mailing list