For our Mac OS run I was able to work with Apple to get performance
numbers up. Mac OS started much, much slower than Linux and failed to
scale at all. One of the Apple engineers emailed with several
recommendations to improve performance. The biggest was to not use
Mpich 1. Once I compiled against Mpich 2 everything went much faster
on the Apple side of things.
Anyways, the Apple run of HPL was compiled against the Accelerate
framework while the Gentoo side was compiled against the most recent
Goto libraries (in January 2006). The Apple side was compiled with a
custom version of GCC 4.0.2 (XCode doesn't come with fortran) and the
Linux side was compiled with GCC 4.0.2 from Gentoo's ebuild.
I may have made it harder for myself in that I compiled a 64 bit
kernel and 64 bit platform.
Our setup is XServe G5's 2.3GHz with 4GB of RAM, running form a local
80GB SATA hard drive. At the time we just had a simple GigE
connection between nodes. Not we have a GigE based FNN.
What environment was your partner using?
On Apr 12, 2006, at 8:26 PM, Justin Bronder wrote:
We've actually noticed the opposite results here with OS X 10.4 vs.
Linux as far as HPL goes.
I don't have the numbers on hand, but Linux was notably faster with
HPL
over both GigE and
Myrinet. In fact the numbers for Colsa (our partner) in the Top500
were
achieved using Yellow
Dog instead of OS X. That being said if you have a moment I have a
few
questions:
1.) What libraries were you using? Did you use Goto's?
2.) Which compiler(s) did you use?
3.) What sort of setup do you have? Diskless or not and so on.
Brady Catherman wrote:
Unfortunately BPS has an annoying habit of overwriting the test
output
files every time it is run regardless of which test you actually run.
I didn't notice this when I was doing my initial stuff on RedHat so I
lost all the other tests results from BPS. I have better tests of the
hardware since then. We ran HPL on the PowerPC cluster running both
Gentoo 2005.1 and Mac OS 10.4 (mac os was ~ %5 faster on average). I
ran IOZone on our XRaid using HFS+ under Mac OS, and Reiser, xfs,
jfs,
and ext3 under Gentoo for PPC64.
I also ran tests using Mr Bayes, ClustalW, MPIBench, and several
other
in house tests that I had available but all of these tests where
either inconclusive or failed to demonstrate any difference between
the platforms.
I will be doing a bunch of this again here in a bit in order to get
data on our new OS build we are putting into production =)
On Apr 11, 2006, at 6:20 PM, Brandon Edens wrote:
Yeah, from the numbers it looks as if it would be dependent on the
purpose of the cluster whether OS X or Gentoo would do better.
On ppc,
Gentoo does poorly on the first two benchmarks and also on context
switching. On x86, the first two are more comparable with RH,
but the
others, Gentoo has a small to large advantage over RH, just as
on ppc.
First and second benchmarks are dhrystone and whetstone, synthetic
benchmarks.
I'd be wary of these types of benchmarks; they've been cheated
before. I'd like
to see some real-world computation benchmarks.
Brandon Edens
[EMAIL PROTECTED] mailing list
--
Justin Bronder
University of Maine, Orono
Advanced Computing Research Lab
20 Godfrey Dr
Orono, ME 04473
www.clusters.umaine.edu
Mathematics Department
425 Neville Hall
Orono, ME 04469
--
gentoo-cluster@gentoo.org mailing list
--
gentoo-cluster@gentoo.org mailing list
