[Beowulf] AMD64 results...

[Bill Broadley]

> and important to some people.  but almost any real code will take
> advantage of having twice as many registers (integer and SIMD).

Indeed, assuming you have the source to recompile.

> don't forget to turn on node interleave in the bios, too.

Assuming a 2.4 kernel I believe that helps single processes running 
on a dual, but doesn't help when 2 processes are running.  With 2.6 I think
it's usually faster to have node interleaving off with 2 processes (if not
1).
> 
> > Kudos for the pathscale-1.4 compiler with -O3.
> 
> ironically, icc -xW generates pretty good-for-opteron code,
> though of course, it's 32b.  I haven't tried using icc to
> generate em64t/and64 code.

With 8.0 and 8.1 I can't seem to get it working on a RHEL x86-64 box
or a Rocks (RHEL-3 based) x86-64 box.

I'll retry with a nacoma based machine and see if I can get the intel
compiler working.

I get errors with 8.0 like:
/usr/bin/ld: skipping incompatible /opt/intel_cc_80/lib/libsvml.a when searching  for -lsvml
or with 8.1:
ld: skipping incompatible /usr/lib64/libm.so when searching for -lm

-- 
Bill Broadley
Computational Science and Engineering
UC Davis