Intel has a bunch of info to aid scientific computational uses, and has
Linux-specific stuff.  My links are not current, but I found this as a clue
or starting point:

http://www.xlsoft.com/en/products/development/intel/mkl.html

That link mentions the following:


The IntelR Math Kernel Library provides developers of scientific and
engineering software with a set of linear algebra, fast Fourier transforms
and vector math functions optimized for the latest Intel PentiumR 4 and
Intel ItaniumR processors. MKL contains LAPACK, the basic linear algebra
subprograms (BLAS), and the extended BLAS (sparse). In addition, there is a
set of fast Fourier transforms in single- and double-precision, real and
complex data types with both Fortran and C interfaces. MKL also includes a
set of vectorized transcendental functions in the Vector Math Library (VML),
offering both high performance and excellent accuracy compared to the libm
functions for most of the processors. MKL is available for the Microsoft
Windows and Linux operating systems.

      Intel Math Kernel Library 5.2 for Windows

      Intel Math Kernel Library 5.1 for Linux


You may be able to locate the right kernel and library links using this and
Munir's info as clues.  Munir's info may already include this kind of thing,
but I think there should be some words about couth with respect to Intel's
recommendations for scientific applications.  The tie between Intel and Red
Hat was/is very strong so it should show when relevant for suitably astute
scientific performance.

Note that it's possible that your applications code was optimized for that
developer's idea (or ignorance) of a host platform, and might even degrade
on a better one unless re-targeted and recompiled.  It's also possible that
a Linux kernel could be optimized for a math-intensive uses and thus get
really big improvements (eg, non-interruptible math runs, etc).  A special
hardware co-processor board might be a productive route also.  Getting away
from multi-user contexts and away from von-Neumann uniprocessors can easily
get 1000-fold improvements (eg, hardware butterfly operators, etc), but
"commercial code" is your big bottleneck.  Those authors may have great
suggestions - if they can "spell" kernel  :-)   Intel and/or your code
developers should have benchmark suggestions.

Happy hunting!

Chuck



> -----Original Message-----
> From: tclug-list-admin at mn-linux.org
> [mailto:tclug-list-admin at mn-linux.org]On Behalf Of Munir Nassar
> Sent: Sunday, October 06, 2002 7:11 PM
> To: tclug-list at mn-linux.org
> Subject: Re: [TCLUG] Kernel - AMD - Computational Speed?
>
>
> On Sun, 6 Oct 2002, Randy Clarksean wrote:
>
> > I do realize that there are operating system issues, etc.
> ... but with all of that I had anticipated a larger reduction
> in computational time.  The code I am running is a commercial
> code that is developed to run on both operating systems, so I
> am fairly sure they work to get the best CPU time on both
> platforms.  I was hoping for something on the order of 1/2
> the CPU time - on large computational runs like this every
> little reduction in time helps.
> >
> > Not being a "kernel" expert by any means ... would it make
> any sense to
> > recompile the kernel on my new platform, rather than relying on the
> > kernel as loaded from the RH7.2 distribution CDs?  Any
> suggestions would
> > be greatly appreciated!
>
> first of all, you should not be using the stock kernels that
> came on CD,
> there were at least 3 upgrades to that kernel,
>
> that said here are some reasons:
> 1. there were some vulnerabilities in the older kernels
> 2. the kernels that come off of the CD by default are
> compiled for i386,
> if you upgrade to something more suitable, like the athlon
> specific kernel
> you will see a very noticable improvement
>
> so go to ftp://updates.redhat.com/7.2/en/os/athlon/ and get
> the athlon
> kernel
>
> and while you are at it, get all the updates from the i686
> area (glibc,
> openssl) and install those too,
>
> once that is done you can install the updates from the i386 area.
>
> Munir Nassar
> RedConcepts.NET
>
> _______________________________________________