On Mon, 2004-05-31 at 10:10, Jakub Jelinek wrote:
> On Mon, May 31, 2004 at 04:50:59PM -0400, Balint Cristian wrote:
> > > Starting with FC3, .i386.rpm and .i686.rpm packages will be compiled
> > > with -march{3,6}86 -mtune=pentium4, so although they will run on
> > > i386 (resp. i686), they will be optimized for P4 (Athlons and newer AMD
> > > chips run P4 optimized code without noticeable performance hit).
> >
> > Is there difference betwwen -mtune=pentium4 and -mtune=athlon ?
> > Pentium4 sounds intel-ish centric but what about AMD ?.
>
> Yes, there is. The thing is, Intel chips are far more sensitive to
> instruction scheduling than AMD chips.
> So, if you run code tuned for AMD on P4, the performance hit is big,
> if you run P4 optimized code on AMD, the performance hit compared to
> running AMD tuned code on AMD is really small.
>
> Jakub
I think it is only fair that someone quantify the performance hit that
K7 and 32bit K8 users can expect versus the improvement that P4 users
will see before this is implemented. Is this change in anticipation of
the Prescott with its even deeper instruction pipeline? It is my
understanding that 32bit-only K8s are shipping already...since they are
being pitch at all areas outside the NA and Europe, they ought to be
considered also before a final decision is made here.
If I am reading this right, not only the kernel and glibc will be
affected...it was stated that the *entire* distro will be optimized for
the P4. People seem to have accepted the idea that K7 optimized kernels
aren't worth the effort...but if we are now talking a performance hit
across the board on all apps not run on Intel hardware, there may be
some unhappy campers. With the 7% performance hit once mentioned here
due to SELinux, just what is a "really small" performance hit? I'm not
sure I will happily accept another 5% loss because I don't run P4s.
Rolling my own kernel is one thing...having to rebuild everything on all
my boxes isn't...I might as well go back to LFS or Gentoo if this is the
case.
It would be interesting to know the ratio of P4 to K7 Fedora users to
determine if this optimization is a great idea since this thread implies
that it would require more resources that RH is willing to apply to also
give us K7 optimized RPMs...or is the logic because corporate America
continues to be sucked in by Intel's marketing? While I have no doubt
that Intel dominates the RHEL market, I do wonder if the same can be
said for the Fedora community.
KAS
