View previous topic :: View next topic |
Author |
Message |
enesideme n00b
Joined: 20 Apr 2002 Posts: 46 Location: france
|
Posted: Fri Jun 14, 2002 7:54 pm Post subject: about gcc3.1 benchmarking |
|
|
I do not know what you aim at when building and building again Gentoo. It may be just for fun. As for me I want a linux-sytem as fast and stable as possible. And of course I want to know if gcc3.1 does really increase performances.
So I try to benchmark gcc3.1-specific optimizations. I compiled two gentoo-systems to do comparative benchmarking. It makes of course no sense to compare what is incomparable. My duron 800 for instance and your PIV or Atlon-XP.
The best to do is to build two or more systems on the same machine with various flags. That's what I have done changing -mcpu=i686 to march=athlon to know if -march=athlon does really improve on my system. here are the results of my tests.
I ran each two times on each gentoo(reboot after each test)
benchmark 1: make dep ; make clean ; sync ; time make bzImage
RESULTS:
-mcpu=i686
1) real 13m37.111
user 12m56.996
sys 0m36.682
2) real 13m37.482
user 12m57.191
sys 0m36.682
-march=athlon
1) 13m37.931
12m57.900
0m36.780
2) 13m38.940
12m57.576
0m36.618
benchmark 2: time tar -jcPf qt-copy-3.0.4.tar.bz2 qt-copy-3.0.4(59.1MB)
-mcpu=i686
1) real 1m51.557
user 1m21.845
sys 0m11.812
2) 1m51.774
1m21.874
0m11.814
-march=athlon
1) 1m52.208
1m21.466
0m11.974
2) 1m52.271
1m21.421
0m11.768
I would not infer too much from these results but I can say with no doubt that the differences are negligible. -march=athlon or -mcpu=i686 it does not seem to make any significant differences for my duron. But I would not infer it never makes any differences. So I would suggest if you have time to spend and room on your disks that you run some comparative benchmarkings. It could help the forthcoming gentoo-users to set the correct values for their flags and avoid them wasting their time. |
|
Back to top |
|
|
squinty n00b
Joined: 20 May 2002 Posts: 5 Location: Kansas City, MO
|
Posted: Sat Jun 15, 2002 12:52 am Post subject: Re: about gcc3.1 benchmarking |
|
|
enesideme wrote: | I do not know what you aim at when building and building again Gentoo. It may be just for fun. As for me I want a linux-sytem as fast and stable as possible. And of course I want to know if gcc3.1 does really increase performances.
So I try to benchmark gcc3.1-specific optimizations. I compiled two gentoo-systems to do comparative benchmarking. It makes of course no sense to compare what is incomparable. My duron 800 for instance and your PIV or Atlon-XP.
The best to do is to build two or more systems on the same machine with various flags. That's what I have done changing -mcpu=i686 to march=athlon to know if -march=athlon does really improve on my system. here are the results of my tests.
I ran each two times on each gentoo(reboot after each test)
benchmark 1: make dep ; make clean ; sync ; time make bzImage
RESULTS:
-mcpu=i686
1) real 13m37.111
user 12m56.996
sys 0m36.682
2) real 13m37.482
user 12m57.191
sys 0m36.682
-march=athlon
1) 13m37.931
12m57.900
0m36.780
2) 13m38.940
12m57.576
0m36.618
benchmark 2: time tar -jcPf qt-copy-3.0.4.tar.bz2 qt-copy-3.0.4(59.1MB)
-mcpu=i686
1) real 1m51.557
user 1m21.845
sys 0m11.812
2) 1m51.774
1m21.874
0m11.814
-march=athlon
1) 1m52.208
1m21.466
0m11.974
2) 1m52.271
1m21.421
0m11.768
I would not infer too much from these results but I can say with no doubt that the differences are negligible. -march=athlon or -mcpu=i686 it does not seem to make any significant differences for my duron. But I would not infer it never makes any differences. So I would suggest if you have time to spend and room on your disks that you run some comparative benchmarkings. It could help the forthcoming gentoo-users to set the correct values for their flags and avoid them wasting their time. |
Um, the -march/-mcpu flags don't necessarily make compile times faster. If you want to compile faster you can use 'make -j<#>' to run make jobs simultaneously and possibly speed up the process. Put simply, what the -march/-mcpu flags are designed to do is optimize the resultant binaries for your architecture, so that your compiled programs will (hopefully) execute faster - a much different benchmark. |
|
Back to top |
|
|
enesideme n00b
Joined: 20 Apr 2002 Posts: 46 Location: france
|
Posted: Sat Jun 15, 2002 9:39 am Post subject: |
|
|
I know but I do not aim at compiling faster. I just aimed at comparing -march=athlon and -mcpu=i686. I just think that -march=athlon is not a better value than -mcpu=i686 for a duron 800. It would be interesting to compare gcc3.1-specific optimization flags for athlon, athlon-Xp, PIV...wi th -mcpu=i686 for instance. I just think my duron 800 does not reap advantage from the value -march=athlon. Binaries don't run faster.
the first test compares compilation time for the kernel the second the time to compress with tar and bzip2. |
|
Back to top |
|
|
dook43 Tux's lil' helper
Joined: 11 Jun 2002 Posts: 116 Location: Baton Rouge, LA
|
Posted: Sat Jun 15, 2002 3:56 pm Post subject: benchmarks |
|
|
Why don't you, instead of running benchmarks that are HDD subsystem dependent (compression/decompression), run SPECint2000 benchmarks under two different kernels or something similar..which will compare apples to apples instead of apples to oranges. _________________ "We who are about to die salute you!" |
|
Back to top |
|
|
enesideme n00b
Joined: 20 Apr 2002 Posts: 46 Location: france
|
Posted: Sat Jun 15, 2002 7:34 pm Post subject: |
|
|
I would like run SPECint2000 but I don't think it is an opensource software and don't want to pay for it. You're right when you say the benchmarks I ran are HDD dependant I mean all tests are anyway hardware dependant but it doesn't really matter since I run my tests on the same box. The only things I change are the values of the compilation flags. I do know no other way to isolate the real effect of a variable. But I can admit the tests I ran may be not the most relevant ones. I would appreciate any suggestions. And maybe could we set something like an experimental procedure so that we could evaluate the real improvement of the C(XX)FlAGS' values. I do not know any recent and opensource benchmarks. That's just why I propose comparative benchmarking. I don't want at all compare apples and oranges since I propose to build two or more gentoo on the same box with various flags. You may obviously 'tarbzip' faster or slower qt-copy on your machine and it makes no sense to compare your results and mine. |
|
Back to top |
|
|
vod Tux's lil' helper
Joined: 10 Jun 2002 Posts: 90
|
Posted: Sun Jun 16, 2002 9:11 am Post subject: you know... |
|
|
that by default the ebuilds for gcc take OUT the CFLAGS you're using right? so no matter what CFLAGS you use, gcc isn't using them. |
|
Back to top |
|
|
enesideme n00b
Joined: 20 Apr 2002 Posts: 46 Location: france
|
Posted: Sun Jun 16, 2002 11:55 am Post subject: |
|
|
you're right for gcc. actually benchmarking kernel compilation benchmarks gcc. So the test I ran is not relevant. So I ran another test. wav to mp3 encoding with gogo since gogo is optimized for mmx 3dnow sse.
time gogo -q 0 Track01.wav 01.mp3
Track01=262,7MB (track one of the wonderfull Koln concert of keith Jarrett)
with -mcpu=i686
1) real 2m57.346
user 2m14.887
sys 0m26.07
2) 2m57.574
2m14.771
0m26.367
with -march=atlon
1) 2m55.531
2m13.189
0m25.723
2) 2m55.747
2m13.046
0m25.723
gogo seems to encode faster with march=athlon for my duron but 1or 2 s /267MB is it much or not? I dont know. Maybe the differences are due (partially or entirely) to my HDD. The two gentoo are not on the same partition and of course the read/write access may be different |
|
Back to top |
|
|
iluvatar n00b
Joined: 16 Jun 2002 Posts: 13
|
Posted: Mon Jun 17, 2002 1:11 am Post subject: |
|
|
Benchmarks are worthless unless you compile glibc, gcc, binutils etc with gcc3.1 and then if you will also recompile those, THEN compiling apps that might actually take serious advantage of optimizations and running those.
Also I don't see ANY point in ANY of these tests? :P _________________ Corona Extra is your friend. |
|
Back to top |
|
|
enesideme n00b
Joined: 20 Apr 2002 Posts: 46 Location: france
|
Posted: Mon Jun 17, 2002 9:45 pm Post subject: |
|
|
I compiled everything with gcc3.1 and the flags -march=athlon or -mcpu=i686 and -fomit frame-pointer -funroll-loops -pipe -03 -m3dnow -mmx for my duron800 even qt to run kde. I just can draw the conclusion that my duron which is an Athlon doesn't take advantage of -march=athlon. I mean my too systems compiled (the first with -march=athlon the second with -mcpu=i686) with gcc3.1 are as fast and stable no more no less. I would have liked to know if an athlon thunderbird or XP and a PIV could take more advantages of the gcc3.1-specific optimization flags for Athlon (Thunderbird XP) or for PIV. Just to know. Anyway I do not plan to buy an thunderbird or Xp or PIV. I would better wait for the soon forthcomming Athlon-64bits. So It does not worry me if nobody wants to waste his time to test the real effect of gcc3.1-specific optimizations for microprocessors. I would add that I appreciate any critical remarks or any suggestions but I haven't red any
constuctive remark or suggestion which could help me to run more relevant benchmarks. But don't worry about it I will keep on searching and sleep soundly while my machine keeps on working |
|
Back to top |
|
|
dook43 Tux's lil' helper
Joined: 11 Jun 2002 Posts: 116 Location: Baton Rouge, LA
|
Posted: Sat Jun 22, 2002 8:12 pm Post subject: benchmarking |
|
|
You should compile things with athlon-tbird. athlon is used only for the old athlon classic's with half-1/3 speed 512K of L2 cache. Your duron will work better with -march=athlon-tbird. _________________ "We who are about to die salute you!" |
|
Back to top |
|
|
|