[Beowulf] AMD64 results...

Robert G. Brown rgb at phy.duke.edu
Sun Dec 19 07:26:06 PST 2004

> > Not sure why the timer is so lousy,
> > I had to make the array large to get a reasonably accurate time:
> This is indeed another interesting point. I'd really like to understand it.
> In addition when I re-run stream the rates vary quite a bit despite 
> the high loop count (50) and very small std dev (min & max are pretty close).

One useful question is what kernel are you running?  Running (e.g.) FC 1
and the 2.4 kernel one would expect to get a wide range of times because
that kernel was more or less broken on X86_64 duals.  I would see vastly
different times for the same job run in multiple instances on the
system, which would certainly increase the spread.

I haven't observed this at all on FC2 (2.6 kernels) and so far FC3 is
behaving itself on the AMD64.

I've got a nasty cold and will probably stay tucked up with my laptop at
least part of the day (it generates enough heat to function as a sort of
hot-water bottle:-).  I may try to run a couple of application-level
benchmarks on metatron to get a rough idea of timings in mixed code.  My
microbenchmark results suggest that it will do pretty well in my Monte
Carlo program because it involves sqrt, ln, exp and some trig per site
in a lattice problem and is hence heavily CPU bound (as opposed to
memory).  If the CPU really does these things 2x-3x faster than a P4 AND
has a higher clock than most of what I have available at home or at
work, I will a happy camper be.

BTW, anyone who tried to visit the benchmaster site I posted earlier and
found the download link broken -- this was pointed out offline and it is
all fixed now.  I'm working on making all the stuff I build and maintain
yum-downloadable, and my onsite php functions had to be hacked to
support the requisite repository structure.  There are binary rpm's for
centos 3.3 (i386 only, which will probably work for RHEL even without a
rebuild), FC2, FC3, and RH 9 (i386 only), as well as source RPMs and
tarballs to match.

I'd be very interested in feedback on how these work as straight RPMs.
I've tested a download/install on a few of my own systems and they seem
to work for me.  Of course, the real fun in having a benchmarking
program is being able to recompile it and play with things, but there is
also at least some virtue in running the exact same code on multiple


Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu

More information about the Beowulf mailing list