[m-rev.] Re: for review: configure option for LARGE_CONFIG

Mark Brown mark at cs.mu.OZ.AU
Tue Feb 14 02:36:53 AEDT 2006


On 13-Feb-2006, Mark Brown <mark at cs.mu.oz.au> wrote:
> Estimated hours taken: 1.5
> Branches: main
> 
> Add a configure option to enable LARGE_CONFIG when compiling boehm_gc.
> This should probably be made the default, but I'll leave that until after
> the performance impact has been measured.

I've done some benchmarks.  The numbers listed here are percentage speedups
(negative values indicate slowdowns).

			saturn		jupiter		pelgrane
			------		-------		--------

speedtest (-O5)		3%		1%		---
speedtest (-O2)		---		---		0.4%
250-queens		-0.9%		2.9%		1.3%
500-queens		12%		1.2%		1.6%
750-queens		-0.8%		---		---

"speedtest" is the script in mercury/tools, run with the default options.
The optimization level in brackets is what the compiler itself was compiled
with, not what it was compiling the benchmark modules with.

The N-queens benchmarks are from the queens_ff module in the tests directory
of the g12 FD library.  That really is a 12% speedup for 500-queens on
saturn; it's not a typo.

"pelgrane" is my laptop: 1.5GHz PPC, 768Mb RAM, Mac OS X 10.4.3.

The increase in executable size is negligible.

These results indicate that there is no reason not to make LARGE_CONFIG the
default.  Next time we import a new upstream Boehm GC, however, we should
run the benchmarks again, since with 4Gb+ RAM becoming more common I wouldn't
be surprised if the defaults were changed to be something more like the
LARGE_CONFIG settings are now.

Any objections to making LARGE_CONFIG the default?

Cheers,
Mark.

--------------------------------------------------------------------------
mercury-reviews mailing list
post:  mercury-reviews at cs.mu.oz.au
administrative address: owner-mercury-reviews at cs.mu.oz.au
unsubscribe: Address: mercury-reviews-request at cs.mu.oz.au Message: unsubscribe
subscribe:   Address: mercury-reviews-request at cs.mu.oz.au Message: subscribe
--------------------------------------------------------------------------



More information about the reviews mailing list