GCC performance testing request

[Thev00d00]

We all love ricer flags, ask any Gentoo user about how much faster they make your system. =O

But seriously, I would be interested in the difference the Graphite framework makes in the newer (4.4+) GCCs with multicore systems.

[Smorg]

Currently on my i7 920 box with gcc-4.4.2

CFLAGS="-march=native -O3 -msse4 -mmmx -floop-interchange -floop-strip-mine -floop-block -pipe"
CXXFLAGS="${CFLAGS}"

With the big 8mb shared L3 cache I think the O3 is probably beneficial more often than not. The last benchmark I read using core2s showed O2 and O3 to be about even with O3 maybe pulling a bit ahead.

[hmmm]

I thought I came across intel doing some gcc compile testing in the latest linux kernel podcast - heres the link http://patchwork.kernel.org/patch/63078/

[birdie]

I thought I came across intel doing some gcc compile testing in the latest linux kernel podcast - heres the link http://patchwork.kernel.org/patch/63078/

The net result doesn't seem obvious to me

The first post clearly shows that -O2 is a winner over -Os, but later the only test won for -O2 kernel and it seemed like given more time and diligence, kernel built with -Os would equal to -O2.

However I digress. I still want someone to run Phoronix tests with these three compilers.

[sabriah]

Can you PLEASE test these versions of GCC: 4.2.4, 4.3.4 and 4.4.2 using whatever benchmarks you like (the more the better).

There was a gcc comparison earlier this year http://www.phoronix.com/scan.php?pag...udio_gcc&num=1


As for the compiler flag choices, check out ACOVEA and http://www.phoronix.com/forums/showp...44&postcount=3


An article by Dunlop and others from 2008, "On the Use of a Genetic Algorithm in High Performance Computer Benchmark Tuning", concluded:

This paper has addressed the issue of extracting the best adapted parameters for the HPL reference benchmark. Adjustment
of the seventeen tuning parameters to achieve maximum performance is a time-consuming task that must be performed
by hand. The use of a genetic algorithm is proposed here to manage this task with individuals corresponding to an
HPL run. Indeed we do not provide here a description of a particular version of a GA. The Acovea framework has been
used to validate the approach over a Beowulf cluster composed of heterogeneous resources: a majority of so-called
“small” nodes and two “large” nodes. In particular, starting from a hand-tuned performance of 84 Gflops, it was possible
to attain the peak performance of 111.6 Gflops on the cluster using a set of parameters determined nearly automatically by
Acovea.

Cheers