Re: Scientific Linux Workstation

From: Pedro Miguel Carvalho (Pedro_MCLX_at_sapo.pt)
Date: 07/01/04 

Date: Thu, 1 Jul 2004 01:31:06 +0100

I have made some simulations for particle physics so I will share my
experience.

I never used MatLab (I use Fortran and C/C++) but most simulations require
tremendous amounts of calculation so the more CPU power the better.
Most simulation can easely be run in multiple processors and "suck them up
dry". So ask the programmer if he will use parrallel processing.
Huge amounts of memory are a must, swapping is a killer. Fill it with RAM,
to the top please :)
Huge amount of disk are with certainty necessary (I used large amounts of
data, 200 plus files each with 500 MB not counting the working data).
A backup system should be in place and be easy to use, even DVD-RAM are a
pain for this much data.
Data movement are probably not the problem since the bottleneck is the
CPU's. In fact I used a NFS file server on a regular 100 Mbps Ethernet
network to move the data around (I was using a distributed system, 6
computers where crunching numbers). A separate thread almost always (99.99%
plus) got the data before it was needed. That said, local storage should be
fast if local working data does not fit in memory. But this is the smallest
of your problems and if your $$$$$ are getting short this is where you can
start to cut.
A UPS might be a good idea if the simulations will run for many days. The
UPS just needs to hold the system until it as saved the working data and
that is just a few minutes, so nothing expensive is needed. Good failsafe
cooling system again if the simulations are long.
About the processor type, my advice is to make a test. Ask the programmer to
make a small test program and then benchmark! If the programmer doesn't
shown interest tell him/her that for a particular program diferent
processors with similar ratings for "normal" programs can have diferences
with a factor of 2 or more for a particular aplication. A few years ago (2
or 3?) I run a program I created in both a Intel Pentium III and a Athalon
(don't remeber the type) with about the same rating for Quake III and lame
mp3 encoder and the Pentium was two time slower than the Athalon (I was
impressed and surprised) and to this day I don't know why.

This are my thoughts. I hope I was of some help.
Greetings,
Pedro Carvalho

Relevant Pages

  • Re: How to get into Scientific Programming
    ... >>simulations of systems, or it could be writing the data acquisition ... were to pass on a Monte Carlo simulation to the wiz programmer, ... the data acquisition system the programmer would have to understand the ...
    (comp.programming)
  • Where to start learning AI
    ... I am a programmer writing the platform for a series of educational ... pair alternatives with each other and ... simulations with a given set of rules for success. ...
    (comp.ai)