Re: NEC-LIST: NEC4 Optimum Desktop Platform

> Hello, everyone. For about five years now I have run f77-compiled
> NEC4 on an HP Apollo 755 workstation. I am looking to replace this
> workstation with newer desktop hardware and would like to have the
> capability to run a 20,000 (twenty thousand) segment (assuming some
> symmetry) problem in less than 24 hours on equipment dedicated to
> this task. Some specs on on my HP 755:
>
> --99 MHz PA-RISC 7100 Processor (124 MIPS, 147 SPECmarks, 40 MFLOPS)
>
> --768 MB total RAM, 256 KB instruction cache, 256 KB data cache
>
> I would appreciate the benefit of any experiences that you can share
> regarding the above and your recommendations of platform, OS, and f77
> compiler. I'd like to keep the total cost around $10k-$15k. I'm
> most interested in computational speed and numerical accuracy.
> Thanks for your time and any comment. Sincerely,

If you have an HP workstation you might as well stick to one.

HP's own propaganda says their processor is right now faster than the
previous benchmark, the Alpha by Digital/Compaq. I don't have any
accurate numbers on that, so this is just usual vendor's tail.

What we have is an C3600 by HP. My NEC compiled with their compiler
and some optimization flags runs a 6493 segments case in 387 seconds
filling and 1470 seconds factor time. If you apply to the nec code
some corrections to make it use lapack, mlib and possibly ATLAS, you
end up with a factor time of about 483 seconds.

Your 20,000 segments scales with 3. So memory for you would be 6.2 GB,
filling time about 1 hour, original factor time 11 hours,
LAPACK/MLIB/ATLAS factor time 3.6 hours. Comes pretty close to your
time requirements.

I can't give you a $$$ on that, I guess German universities pay a bit
different bucks than US government agencies do. We have payed about
28,000 DM which is about 14,000 US$.

There is no way to use any INTEL/AMD computer for this large of a
problem when you stick with NEC. Even Atlas makes the AMD 900 MHz run
a factor of 3 slower than the C3600 (most of the time the AMD is about
5 - 8 times slower). If you were able to plug enough memory in it,
that is.

You might also want to consider getting a newer code than NEC. NEC is
very good, tested, retested and evaluated, hence reliable. Perhaps you
need this.

There are a couple of good commercial codes around. For combining
high frequency approximations (physical optics, UTD,...) with accurate
MoM, look for FEKO (http://www.emss.co.za/, I personally have no
interest in this company I'm just a happy user). Look for CTS
Microwave Studio, http://www.cst.de (I have no experience with that
code), Concept (University of Hamburg-Harburg), and probably others
that I've forgot (look at the ACES website for more information).

cheers
juergen
Received on Fri Feb 09 2001 - 14:29:16 EST