Originally posted by: litfos

Hello all,
I tried to compile apache 2.2.14 first at work on an idle lpar (apache sources were located in /tmp).
aix 5.3 TL 10.

System Model: IBM,8203-E4A
Processor Type: PowerPC_POWER6
Number Of Processors: 4
Processor Clock Speed: 4204 MHz
Kernel Type: 64-bit
Memory Size: 9728 MB

with gcc 4.2,

$ date ; ./configure> /dev/null ;date
Mon Nov 23 16:26:59 CUT 2009
Mon Nov 23 16:31:59 CUT 2009
$ date ; gmake > /dev/null ;date
Mon Nov 23 16:34:57 CUT 2009
Mon Nov 23 16:40:49 CUT 2009

It was successful.
and then did the same at home on a 2.26 Ghz Intel core 2 Duo apple laptop with 10.6.2 OS(darwin), 2 GBs of RAM.
date; ./configure >/dev/null;date
Mon Nov 23 23:56:19 CET 2009
Mon Nov 23 23:56:27 CET 2009

date; make >/dev/null 2>/dev/null ;date
Mon Nov 23 23:57:50 CET 2009
Mon Nov 23 23:58:34 CET 2009
I was surprised that compilation on a laptop was much faster.
Is there an explanation for these differences in timings?
#AIX-Forum

Originally posted by: shargus

Try running your test again, but use the "time" command to time them; i.e.

time ./configure

time gmake

Then compare the user and sys times reported, not just the real (or "wall clock") times. That will tell you the amount of time spent actually running your processes.

Your laptop is able to devote all of it's CPU cycles to your compile process, so it completes faster. Your P6 server is probably doing other things before it gets around to running your compile.
#AIX-Forum

Originally posted by: niella

Hi,

The RPM packages like gmake you use on AIX may not have been compiled with POWER6 optimization flags (qtune / qarch), so it is not really a "fair" comparison.

Niel
#AIX-Forum

Originally posted by: jklotz

I found in previous tests that using bash during a configure process makes that one often much faster under AIX, eg :

export CONFIG_SHELL=/opt/freeware/bin/bash

juste before launching the ./configure.

(which implies you installed the bash rpm of course ;) ).
#AIX-Forum

Originally posted by: litfos

Hello All,

thank you for you comments.

As @shargus sugested:
$ time ./configure>/dev/null

real 5m10.85s
user 0m36.72s
sys 0m25.27s
$ time gmake > /dev/null

real 5m56.77s
user 1m59.78s
sys 0m24.70s
And as @jklotz suggested:
$ export CONFIG_SHELL=/opt/freeware/bin/bash
$ time ./configure>/dev/null

real 1m58.61s
user 0m35.81s
sys 0m23.88s

On a Mac:
$ time ./configure >/dev/null

real 0m9.181s
user 0m3.435s
sys 0m4.708s

$ time make >/dev/null
real 0m41.099s
user 0m23.652s
sys 0m20.570s
Few observations, I still see huge difference in 'user" but what surprises me such big boost with using different shell than ksh for configure script.
@niella I guess you suggest to recompile gcc and not gmake as you suggested? Do you know what do I have to pass as configure parameters to get gcc compiled with Power6 optimization?

P.S Do you get on your machines similar results for such compilation?
#AIX-Forum

Originally posted by: jklotz

With one 5000 mhz processor :

time configure --prefix=/opt/httpd

real 1m40.16s
user 0m20.68s
sys 0m15.25s

time make

real 1m16.89s
user 0m35.71s
sys 0m18.56s
This is apache 2.2.14, using the ibm xl c compiler.
My understanding from several findings on the net is that ksh works with some temporary files, which slows it much, however i did not make specific tests to support this.
#AIX-Forum

Originally posted by: jklotz

By the way, I wouldn't use a compilation process as a benchmark reference between platforms, as the compiler has to produce machine code versus a specific processor :)
#AIX-Forum

Originally posted by: litfos

Hello Jklotz,

I agree that compilation is not a best way to check the cpu speed.
I noticed that Phoronix http://www.phoronix.com/scan.php?page=article&item=freebsd8_ubuntu910&num=1 in their benchmarks in one of the steps are converting wav file to mp3. As this is a simple task to run but very CPU intensive I decided to do the same.

I took this wave file from Phoronix website and precompiled lame version:
http://www.phoronix.net/downloads/phoronix-test-suite/benchmark-files/pts-trondheim-wav-3.tar.gz
http://www.perzl.org/aix/index.php?n=Main.Lame

Power6
$ time lame pts-trondheim-3.wav
real 0m50.74s
user 0m50.15s
sys 0m0.11s

And on a Mac

real 0m26.878s
user 0m25.886s
sys 0m0.239s
What I did additionally was starting the same encoding on Power5 machine:
System Model: IBM,9133-55A
Processor Type: PowerPC_POWER5
Number Of Processors: 1
Processor Clock Speed: 1900 MHz
Kernel Type: 64-bit
Memory Size: 3072 MB

real 0m58.99s
user 0m58.77s
sys 0m0.10s

I am wondering how should I read then these 4.2 Ghz in the Power6 box. As the difference between Power5 1900 and Power6 4.2 is not really much visible as I would expect?

Can you suggest any other software which would be good for such CPU benchmarkings?
#AIX-Forum

Originally posted by: jklotz

Roughly, I wouldn't expect x2 performance in favor of the core 2. I'm not sure how to qualify the .wav conversion (int of floating point manipulation), but it doesn't match with the results you can collect here on the spec website :

http://www.spec.org/cpu2006/results/

(expect to have a bit of maths to make to convert multicore-configurations results to one processor; of course the results are approximations, but their order of amplitude is usually in favor of the Power 5 - in my case).
#AIX-Forum

Originally posted by: j.gann

lame also seems not to be a fair benchmark to me because it uses assembly language modules on x86 for optimization. I doubt that given the (low) marketshare of power6 in multimedia applications anyone will ever write these modules in power assembly language.
#AIX-Forum

Originally posted by: litfos

Hi,

today I tested The Fhourstones Benchmark http://homepages.cwi.nl/~tromp/c4/fhour.html
This integer benchmark solves positions in the game of connect-4, as played on a vertical 7x6 board.

I compiled the test program wih gcc and xlc 10. No special difference was observed in timings.
Power5
xlc
$ time ./SearchGame < inputs >/dev/null

real 12m15.67s
user 12m14.27s
sys 0m0.00s
$
gcc
$ time ./SearchGame.2 < inputs >/dev/null

real 12m6.00s
user 12m4.30s
sys 0m0.00s
Power6 4.2 gcc
time ./SearchGame < inputs >/dev/null

real 8m50.46s
user 8m47.31s
sys 0m0.00s
Mac
time ./SearchGame < inputs > /dev/null

real 4m38.767s
user 4m29.344s
sys 0m0.990s
Still Core 2 visibly faster.
I run http://math.nist.gov/scimark2/ on a Mac too:
./scimark2

• **
• SciMark2 Numeric Benchmark, see http://math.nist.gov/scimark **
• for details. (Results can be submitted to pozo@nist.gov) **
• **

Using 2.00 seconds min time per kenel.
Composite Score: 265.12
FFT Mflops: 163.76 (N=1024)
SOR Mflops: 571.61 (100 x 100)
MonteCarlo: Mflops: 59.30
Sparse matmult Mflops: 232.86 (N=1000, nz=5000)
LU Mflops: 298.05 (M=100, N=100)
I would be grateful if someone could post results of this benchmark from power6 5GHz box.
P.S Can anyone recommend me a benchmark which measures memory access speed? I am mainly interested in benchmarking access to the shared memory, attaching and detaching from it.
#AIX-Forum