Re: [ANNOUNCE] ktimers subsystem

[Daniel Walker]

On Sat, 2005-09-24 at 15:56 +0200, Thomas Gleixner wrote:
> On Sat, 2005-09-24 at 12:35 +0200, Roman Zippel wrote:
> > Hi,
> > 
> > On Sat, 24 Sep 2005, Ingo Molnar wrote:
> > 
> > > > Anyway, the biggest cost is the conversion from/to the 64bit ns value 
> > > > [...]
> > > 
> > > Where do you get that notion from? Have you personally measured the 
> > > performance and code size impact of it? If yes, would you mind to share 
> > > the resulting data with us?
> > > 
> > > Our data is that the use of 64-bit nsec_t significantly reduces the size 
> > > of a representative piece of code (object size in bytes):
> > > 
> > >                 AMD64    I386        ARM          PPC32       M68K
> > >    nsec_t_ops   226      284         252          428         206
> > >    timespec_ops 412      324         448          640         342
> > > 
> > > i.e. a ~40% size reduction when going to nsec_t on m68k, in that 
> > > particular function. Even larger, ~45% code size reduction on a true 
> > > 64-bit platform.
> > 
> > Without any source these numbers are not verifiable. You don't even 
> > mention here what that "representative piece of code" is...

These numbers are misleading .. Doing a total code comparison shows that
a 2.6.14-rc2+ktimers kernel is slightly bigger than a vanilla 2.6.14-rc2
kernel (gcc 4.0, defconfig) .. So your argument that "small is faster"
must mean ktimers is slower, or at least not faster ..  

Making a speed argument based on code size doesn't make much sense to
me, if it's actually faster then show that it's faster. 

Daniel


-
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/