On Sun, 23 Jun 2002, Brian Hechinger wrote:

> On Thu, Jun 20, 2002 at 02:34:21PM +0100, David Brownlee wrote:
> >
> > 	When changing the default it will be important to ensure we do
> > 	not impair the binaries performance on other (non supersparc)
> > 	sparc systems. +10% on a supersparc is probably worth -1% on a
> > 	sun4c, but not -10% :)
>
> actually, i wasn't recommending the changing of the default setup, i was
> thinking more along the lines of a list of optimizations per platform.  in
> my case at the very least, i don't mind rebuilding everything, since i usually
> end up doing it anyway, so i just want to be able to pop the correct
> optimization flags in where apropriate.

	-mtune=supersparc may gain significantly on supersparc without
	noticibly affecting performance on pre-supersparc systems.
	If so, it certainly makes sense to change it to be the default.

	Obviously compiling for a specific machine would want -mcpu=XXX
	for appropriate values of XXX...

	Personally I'm much more interested in finding the best default
	settings as that would give the maximum gain for least ongoing
	effort :)

> however, i do see the benefit of pre-built optimized systems for install and
> forget setups.  so we should find the nice happy medium for the default install
> and then maybe offer optimized builds.  which i would be more than happy to do
> once i get things setup here.

	Building a complete distribution optimised for a specific processor
	should be as simple as a single command to invoke build.sh with
	the appropriate arguments.

> > 	Possible (quite ambitious) suggestion:
> > 	    - Build -current system with and without -mtune=supersparc,
> > 	      including kernel & boot blocks.
> > 	    - Compare disk usage (Not important, just interesting number :)
> > 	    - Setup a script to perform some benchmarks - ssh -1, ssh -2,
> > 	      maybe something interesting from pkgsrc/benchmarks.
> > 	    - Compile static version of the test binaries and repeat test
> > 	      to confirm they show the same behaviour as dynamic.
> > 	    - Put up test binaries and benchmark script for people with
> > 	      other sparc models to test and report.
>
> if no-one else does this, i will start doing it after i get my two new mail
> server boxes so i can retire my old mail server to the duty of performance
> testing for port-sparc.

	Great! :)

-- 
		David/absolute		-- www.netbsd.org: No hype required --