Date: Wed, 25 May 2005 19:51:04 +0200 From: =?ISO-8859-1?Q?K=F6vesd=E1n_G=E1bor?= <gabor.kovesdan@t-hosting.hu> To: Nicolas Blais <nb_root@videotron.ca> Cc: freebsd-amd64@freebsd.org Subject: Re: amd64 optimized gcc? Message-ID: <4294BB08.3090703@t-hosting.hu> In-Reply-To: <200505251317.22128.nb_root@videotron.ca> References: <200505251317.22128.nb_root@videotron.ca>
next in thread | previous in thread | raw e-mail | index | archive | help
Hello, As for the optimization, I'm very interested how fast can be an actual optimized system. The stock release doesn't optimizes too much, afaik it uses -O. I have been experimenting for a while with building the whole system with my custom cflags, and yesterday I succeeded. I managed to build the whole system with CFLAGS=-s -Os -march=athlon64 COPTFLAGS=-s -Os -march=athlon64 I might have used -O2 instead of -Os but not -O3 which made a compilation error. The -s is for srtipping the binaries and libs, because this is intended to be used on a production system, where we use stable software releases, thus we don't need debug symbols And I ripped out the softwares I don't need with macros. (named, sendmail, games, ...) Unfortunately I don't have a machine for testing purposes but I would ty it. If I provide You such a disc would You try it in the same way? I'm not sure it is a working a disc, but it should be. I haven't read that anybody made such discs. Cheers, Gábor Kövesdán Nicolas Blais wrote: >I am developping a software that follows a random()-dependant algorithm which >is extremely cpu intensif. >I decided to run on different platforms to see how it performed based on cpu >and os (in a way of benchmarking) and I'm surprised by the numbers: > >Reference times for benchmark (5e+07 run of the algorithm): > >(FreeBSD/i386) Venice (S939, 512K L2 cache) Athlon64 3000 overclocked @ 2655 >Mhz : 78.3072 s (638511 r/s) > Note: Cool 'n' Quiet! Disabled in BIOS. > Note: 1 G RAM > >(FreeBSD/amd64) Venice (S939, 512K L2 cache) Athlon64 3000 overclocked @ 2655 >Mhz : 71.2521 s (701732 r/s) > Note: Cool 'n' Quiet! Disabled in BIOS. > Note: 1 G RAM > >(Knoppix/i386) Clawhammer (S747, 1MB L2 cache) Athlon64 3200 @ 2000 Mhz : >133.858 s (373325 r/s) > Note: Compaq R3240CA Laptop, Cool 'n' Quiet! forced by BIOS. > Note: 512 M RAM > >(FreeBSD/amd64) Clawhammer (S747, 1MB L2 cache) Athlon64 3200 @ 2000 Mhz : >47.2754 s (1057630 r/s) > Note: Compaq R3240CA Laptop, Cool 'n' Quiet! forced by BIOS. > sysctl hw.acpi.cpu.px_control=-1 > Note: 512 M RAM > >(FreeBSD/i386) Pentium II @ 233 Mhz : 538.136 s (92913.3 r/s) > Note: 192 M RAM > >Not surprising is the Pentium II :). What is surprising is that amd64 FreeBSD >seems to execute code faster than i386 FreeBSD, so I'm wondering if gcc >(amd64) really optimizes code for the cpu. If it is, I would probably move my >httpd server to amd64... > >Also, maybe less surprising is that Knoppix sucks running the algorithm for >some reason and that L2 cache really is a big factor (my Laptop outperforms >my heavily overclocked box). > >Any comments? > >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4294BB08.3090703>