[Beowulf] [AMD64] Gentoo or Fedora
ed at eh3.com
Fri Aug 31 12:54:40 PDT 2007
On Fri, 31 Aug 2007 12:56:48 -0400 Larry Stewart wrote:
> For X86 archtectures, it actually makes a very noticable performance
> change (30%?) to use build switches that are appropriate for your
> machines, rather than the generic switches chosen by Red Had or
Before adding to this topic, I'd just like to say that I have nothing
against any of the "build-from-source" distros. There is a wealth of
good information in, for instance, some of the Gentoo forums and I do
appreciate the effort that others have put into them.
Getting back to the topic, I often hear the "if you compile from
scratch with FOO optimizations/settings turned on, then you'll see BAR
improvement" where BAR is typically expressed as some very hazy value
or is merely stated as "a lot!" with overuse of the shift keys and no
benchmarks in sight.
So I'm calling "baloney".
Over the years, I've (re-)compiled a number of applications including
many different types of math/physics models and, when using the *same*
compiler with *typical* distro flags, I have yet to see appreciable
(that is, greater than 10%) improvements. And when I say *typical*
please understand that the default compiler flags for many distros
(e.g., Fedora) are already "-O2 -march=...".
At the same time, I have routinely seen >10% differences (sometimes
much more) when using two or more different compilers on certain
architectures and/or codes (e.g., comparing different versions of GCC or
one version of GCC vs. an Intel compiler). And so have many others:
Yet we don't see folks building distros using, say, the PGI or Intel
or PathScale compilers. If someone is actually doing that (and it
works), then please speak up...
So why are these rumors (that is, the advantage of from-source vs.
pre-built distros) so persistent when its so hard to find any
I have a theory. Call me a cynical BOFH, but I think a lot of folks
get very excited about the prospect of "tweaking" things. They want to
believe that going from -O2 to -O3 will be a big difference because,
well, theirs goes to "11". And they are not interested in rational/
repeatable measurements which will, I'll wager, likely show that their
additional compile times (and effort) will never be won back by the
amortized effect of such minuscule gains.
But maybe I'm just turning into a cynical old man. Maybe I should
abandon rational thought and concentrate on having fun pimping my
ps - Please, if you are going to take issue with this post then
by all means bring with you some actual benchmark numbers
that we can reproduce on our laptops and desktops. Bonus
points if you can write a script that does a semi-decent
job of automating such a benchmark.
Edward H. Hill III, PhD | ed at eh3.com | http://eh3.com/
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 189 bytes
Desc: not available
Url : http://www.scyld.com/pipermail/beowulf/attachments/20070831/f43aabe3/signature.bin
More information about the Beowulf