[Warewulf] Re: [Beowulf] hpl size problems

Greg M. Kurtzer gmk at runlevelzero.net
Mon Sep 26 13:18:03 PDT 2005

On Mon, Sep 26, 2005 at 04:01:06PM -0400, Andrew Piskorski wrote:
> On Mon, Sep 26, 2005 at 03:43:00PM -0400, Andrew Piskorski wrote:
> > It's amusing that Mark Hahn is already participating in this thread,
> > because his post to the Beowulf list gave a link explaining a detailed
> > real-world example of that effect very nicely:
> > 
> >   http://www.beowulf.org/archive/2005-July/013215.html
> >   http://www.sc-conference.org/sc2003/paperpdfs/pap301.pdf
> Hm, looking at it again reminds me that Petrini, Kerbyson, and Pakin's
> paper is much more general than I implied.  E.g., their "Finding #4"
> is:
>   "Substantial performance loss occurs when an application resonates
>   with system noise: high-frequency, fine-grained noise affects only
>   fine-grained applications; low-frequency, coarse-grained noise
>   affects only coarse-grained applications."
> They also mention that LINPACK is coarse-grained with not that much
> communication, so I guess my naive assumptions about HPL were wrong.
> Hm, no doubt Mark H. knew that, which is why he didn't immediately
> jump on the "Warewulf is faster because it eliminates the
> jitter-causing daemons" hypothesis.

I snipped out a bit of content above which may have led to a
misunderstanding... The application that got the 30% speed up was not
the HPL, rather a geophysics application (developed in house) which
IIRC does barrier quite a bit and is high-frequency/fine-grained.
Your assumption seems accurate indeed. ;)

Actually, our HPL speedup was not as significant. The fastest previous
configuration (as Dell mentioned to us) was 81%, and we got 83.4%. With
that said, our system had faster CPU's but the same bus speeds as the
previous record holder which even makes it more remarkable that we
obtained a higher efficiency.

Thanks for the link to the PDF!

Greg M. Kurtzer

Do not look anywhere for truth, for all that is needed is to refrain from
allowing concepts to arise.

More information about the Beowulf mailing list