gcc-3.3.6 vs gcc-4.2.2 scimark performance degradation

gcc-3.3.6 vs gcc-4.2.2 scimark performance degradation

[CH Gowri Kumar - గౌరీ కుమార్]

Hi all,

We are observing a slowdown in the performance of scimark benchmark
w.r.t to the gcc versions 3.3.6 vs 4.2.2. And in our actual application, we
are observing a slowdown of almost 20%..

We are currently using -O2 for compilation for both the gcc versions.
On our product, even with -O3 on gcc-4.2.2, we are not able to reach the
performance of gcc-3.3.6

We are also observing that the size of binary has increased  from 38M to 50MB.
Compiling with -Os didn't yield much benefit (resultant binary was 47 MB).
We were guessing that the increase in size could probably because
of inlining and hence should yield a better performance. But the results
are  disappointing.

We have done the profiling of the application code using gpror and we
observe that
almost 80% of the functions of gcc-4.2.2 are running slower than the
ones on gcc-3.3.6

Did anyone face this sort of issue earlier (slowness 3.x vs 4.x) ?
Could someone suggest the potential list of flags of 4.2.2, which can
probably yield
a better result or at least at par with the 3.3.6, so that we can
experiment a bit
and arrive at the best combination. ( we cannot use the -march flag
and -ffast-math,
due to strict requirement IEEE floating-point conformance )

I have searched through the gcc bugs list, and can find one bug  for
scimark, but
it was not conclusive.(http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33431)


The following are the relavent details of gcc versions and scimark results.
Machine
=======
Linux opteron26 2.6.9-67.ELsmp #1 SMP Wed Nov 7 13:56:44 EST 2007
x86_64 x86_64 x86_64 GNU/Linux

gcc-3.3.6
========
$gcc -v
Reading specs from
/linux/depot/gcc-3.3.6-RHEL_4/bin/../lib/gcc-lib/x86_64-redhat-linux/3.3.6/specs
Configured with: ../../src/gcc-3.3.6/configure
--prefix=/depot/gcc-3.3.6-RHEL_4 --disable-shared --disable-checking
--with-system-zlib --enable-threads=posix --enable-__cxa_atexit
--enable-languages=c,c++,f77,objc --host=x86_64-redhat-linux
Thread model: posix
gcc version 3.3.6

------------------------------------------------------------------------------------
**                                                              **
** 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        :    749.67
FFT                   Mflops:   742.23    (N=1024)
SOR                   Mflops:   648.81    (100 x 100)
MonteCarlo:       Mflops:   317.68
Sparse matmult  Mflops:   942.96    (N=1000, nz=5000)
LU                       Mflops:  1096.65    (M=100, N=100)
======================================================


gcc-4.2.2
=======
gcc -v
Using built-in specs.
Target: x86_64-redhat-linux
Configured with: ../../src/gcc-4.2.2/configure
--prefix=/depot/gcc-4.2.2-static --disable-shared
--enable-threads=posix --disable-checking --with-system-zlib
--enable-__cxa_atexit --disable-libunwind-exceptions
--enable-languages=c,c++,objc,fortran --with-cpu=generic
--host=x86_64-redhat-linux
Thread model: posix
gcc version 4.2.2

**                                                              **
** 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            :   715.93
FFT                  Mflops    :   676.06    (N=1024)
SOR                  Mflops   :   659.92    (100 x 100)
MonteCarlo:       Mflops  :   349.75
Sparse matmult  Mflops  :   801.66    (N=1000, nz=5000)
LU                      Mflops  :  1092.27    (M=100, N=100)
=====================================================

Thanks in advance.

Regards,
Gowri Kumar
www.gowrikumar.com