GCC/CFLAGS Benchmark for HEAD and PHP_5_0 View as Markdown

Hello Sebastian,

could you please repeat the tests for php 4.3 to see the interesting
difference? Also how about providing a spreadsheet for a better overiew?
How did you do the timeings, i assume time command? And what CPU did you
us, especially how much cache arewe talking about?

best regards
marcus

Monday, October 4, 2004, 8:30:47 PM, you wrote:

I ran the Zend/bench.php script with PHP 5.0 and PHP 5.1 compiled with
GCC 3.3 and GCC 3.4 using different optimizations on my Intel Pentium-M
laptop:

HEAD (PHP 5.1.0-dev)

GCC 3.3.4 (-march=pentium3)

  -O0:                          48.602
  -Os:                          29.920
  -O1:                          31.349
  -O2:                          29.029
  -O3:                          29.644

GCC 3.4.2 (-march=pentium-m -mtune=pentium-m)

  -O0:                          48.966
  -Os:                          26.286
  -O1:                          29.253
  -O2:                          29.100
  -O3:                          27.767

  -fprofile-{generate|use} -Os: 24.216
  -fprofile-{generate|use} -O1: 26.575
  -fprofile-{generate|use} -O2: 26.339
  -fprofile-{generate|use} -O3: 25.537

PHP_5_0 (PHP 5.0.3-dev)

GCC 3.3.4 (-march=pentium3)

  -O0:                          58.394
  -Os:                          42.570
  -O1:                          43.454
  -O2:                          42.092
  -O3:                          42.066

GCC 3.4.2 (-march=pentium-m -mtune=pentium-m)

  -O0:                          59.272
  -Os:                          35.853
  -O1:                          38.275
  -O2:                          37.989
  -O3:                          41.020

  -fprofile-{generate|use} -Os: 33.926
  -fprofile-{generate|use} -O1: 36.853
  -fprofile-{generate|use} -O2: 35.335
  -fprofile-{generate|use} -O3: 38.897

For the -fprofile-{generate|use} builds I used the Zend/bench.php
script, too, to generate the profile information. If we were to add
Makefile target (like GCC itself ("make profiled-bootstrap")) to the
PHP build that makes use of -fprofile-{generate|use} it would make
sense to use "make test" to generate the profiling information.

When I find the time I will repeat the benchmark on my AMD Athlon64 box
because I think the small L1/L2 caches of the laptop CPU affect the
performance of the bigger code generated through optimizations. You can
clearly see that code optimized for size (-Os) is faster than the more
aggressive -O1, -O2, and -O3 optimization levels.

--
Sebastian Bergmann http://www.sebastian-bergmann.de/
GnuPG Key: 0xB85B5D69 / 27A7 2B14 09E4 98CD 6277 0E5B 6867 C514 B85B 5D69

--
Best regards,
Marcus mailto:helly@php.net

Very interesting numbers, I'd like to second Marcus' request for a 4.3
benchmark.

I was somewhat surprised by O2 and O1 being slower then Os, while O3 in
some cases may end over optimizing which would it explain it's poor
showing. However, it could be that it makes simple situations slower,
while more complex operations that are generally more CPU intensive will
in fact become faster. If you don't mind, could you please all include
data for "time make test" as it seems to cover a much greater quantity
of code.

Ilia

I've done lots of testing with these kind of compiler flags. It's really a
tough call because you might change 2 lines in the PHP code and then your
results might differ completely because of CPU cache hits/misses. Also very
much depends on the CPU.
But in any case, interesting results.

Andi

At 08:30 PM 10/4/2004 +0200, Sebastian Bergmann wrote:

I ran the Zend/bench.php script with PHP 5.0 and PHP 5.1 compiled with
GCC 3.3 and GCC 3.4 using different optimizations on my Intel Pentium-M
laptop:

HEAD (PHP 5.1.0-dev)

GCC 3.3.4 (-march=pentium3)

  -O0:                          48.602
  -Os:                          29.920
  -O1:                          31.349
  -O2:                          29.029
  -O3:                          29.644

GCC 3.4.2 (-march=pentium-m -mtune=pentium-m)

  -O0:                          48.966
  -Os:                          26.286
  -O1:                          29.253
  -O2:                          29.100
  -O3:                          27.767

  -fprofile-{generate|use} -Os: 24.216
  -fprofile-{generate|use} -O1: 26.575
  -fprofile-{generate|use} -O2: 26.339
  -fprofile-{generate|use} -O3: 25.537

PHP_5_0 (PHP 5.0.3-dev)

GCC 3.3.4 (-march=pentium3)

  -O0:                          58.394
  -Os:                          42.570
  -O1:                          43.454
  -O2:                          42.092
  -O3:                          42.066

GCC 3.4.2 (-march=pentium-m -mtune=pentium-m)

  -O0:                          59.272
  -Os:                          35.853
  -O1:                          38.275
  -O2:                          37.989
  -O3:                          41.020

  -fprofile-{generate|use} -Os: 33.926
  -fprofile-{generate|use} -O1: 36.853
  -fprofile-{generate|use} -O2: 35.335
  -fprofile-{generate|use} -O3: 38.897

For the -fprofile-{generate|use} builds I used the Zend/bench.php
script, too, to generate the profile information. If we were to add
Makefile target (like GCC itself ("make profiled-bootstrap")) to the
PHP build that makes use of -fprofile-{generate|use} it would make
sense to use "make test" to generate the profiling information.

When I find the time I will repeat the benchmark on my AMD Athlon64 box
because I think the small L1/L2 caches of the laptop CPU affect the
performance of the bigger code generated through optimizations. You can
clearly see that code optimized for size (-Os) is faster than the more
aggressive -O1, -O2, and -O3 optimization levels.

--
Sebastian Bergmann http://www.sebastian-bergmann.de/
GnuPG Key: 0xB85B5D69 / 27A7 2B14 09E4 98CD 6277 0E5B 6867 C514 B85B 5D69