finding the best optimization flags via genetic algorithm

Fingolfin · n00b Joined: 23 Apr 2002 Posts: 32

Very interesting article here: http://www.coyotegulch.com/acovea/index.html. I found this on www.osnews.com. There is also a recently posted review of Gentoo.

ewan.paton · Posted: Tue Nov 18, 2003 5:44 am Post subject:

genetic algorithms make my head hurt i once atempted to use them for what i thought was going to be some production line optimization stuff, after a week or so the variables just kept growing and growing and i went sod this for a laugh. i was geting to the point of having to do some realy serious maths just to get the most rudimentory results plus i was doing it in my free time so nobody knew i failed (to be fair cusum graphs were new to them so it would have been a nightmare to explain itl)
_________________
Giay tay nam | Giay nam cao cap | Giay luoi

Bytal · Posted: Tue Nov 18, 2003 11:30 pm Post subject:

On one hand this could be interesting for Gentoo in the sense that theoretically, you can have an automated program determine your gcc optimizations for each program. On the other it doesn't even seem feasible to do this for each individual compile because of the large compile times and lack of benchmarking for a lot of desktop and other regular use applications. However, and that's a big however, it might be interesting to see if it would be possible to come up with a standardized benchmarking tool and some sort of distributed test project, where a large number of people would run the automated benchmarking on at least a few of their compiles and report them to a centralized DB. This way we could have a large number of "standardized" benchmark scores for various applications/processor types. Of course at this point I see more problems then solutions besides the obvious facts that people can't even agree on what to optimize for, and the difficulty in coming up with appropriate benchmarking software.
_________________
Of all the things I lost, I miss my mind the most.

ewan.paton · Posted: Wed Nov 19, 2003 6:42 am Post subject:

that would have to be some benchmark tool ignoring the different cpu speeds youd also have to contend with memory latency, disk access time, motherboard chipsets, what daemons are runing and at what load, disk useage, kernel version, nice level and probably a bunch of other stuff[1]. its a nice idea in princaple but these days if you realy need the performance its easyier to put your hand in your pocket[2] if your joe blogs and if you do something wacky like earthquake simulation you benchmark your own software anyway

[1] ever read a pc mag review of 12 different pc's using the same processor, price and made generic hardware mostly useing the same chipsets and see the varying scores
[2] an athlon2600 motherboard, processor, cooler combo will cost under £90 these days flog the old gear and the upgrade should cost under 50 notes, and repeat every 4-6 months or keep your old gear and build a realy ramshackle cluster
_________________
Giay tay nam | Giay nam cao cap | Giay luoi

robmoss · Posted: Sat Nov 22, 2003 2:40 pm Post subject:

I've written an ebuild for this programme, but it's still in its infancy.

Anyone know how I can autodetect the architecture better? Right now there's serious problems if someone uses "-march=arch -mcpu=arch"

The ebuild is currently hosted here:

http://home.jesus.ox.ac.uk/~rmoss/acovea/

EDIT: If you're using an architecture other than Athlon XP at the moment, you need to change the line of the patch that says:

masseya · Posted: Fri Jan 23, 2004 3:52 pm Post subject:

Please follow up on this topic to the newer and more informative thread here.

Moving from Gentoo Chat.
_________________
if i never try anything, i never learn anything..
if i never take a risk, i stay where i am..