RE: LAM/MPI Parallel processing

[Simon Valiquette <Simon.Valiquette@example.com>]

  Thank you to everybody that answered me eiter on or off the mailing
list (including Mr. Turnbull).  I don't have yet the solutions to all my
problems, but at least I know much more where and to who I can ask
questions.  So, I'll send the futurs emails off the mailing list (you'll
have to send me a email if you are really interested).  Just giving
there last answer on the mailing list about my system, if any one
interested.

   Yesterday, I've talked quickly to Mr. Jon Hall about my simulator and
he was kind enought to give me some advices.  Essentially, he told me
that making a good parallel systems is usually not something trivial. 
There is so much differents parameters to care about, that the best way
to build such systems is often just try making the software and building
the cluster of computers in a reasonnable way.  If it work well enought,
you are happy.  If not, then you have to find out why and possibly
redesign the whole thing.  To make good parallel system design, you need
good understanding of the hardware, usually depth understanding of the
problem to resolve, and experience in making parallel computing.  There
is not a lot of ways to get the experience, except making a lot of
mistakes, and learning from them as quickly as possible.

   He also told me to try to use, if possible, only little endian
computers or only big endian computers to avoid the bit order swaping
(but I will still mix both, because the 2 Alpha I have are quite fast
and unused anyway, and that there is many available good PC, but they
are slowers).

   By the way, if I believe his conference, Linux is now more used on
desktop than MacOS.  Interesting information for your ISP if they
support both MAC and Windows, but say that they don't support Linux
because this market is too small for them...

> From: Stephen J. Turnbull [mailto:turnbull@example.com]

> Don't abuse names (or other words, for that matter) only for
> laziness's sake.  Among other things, it makes it much harder to
> receive or give good advice.
> 
>     mumei> Maybe it's just a kind of technical language evolution.
> 
> _De_volution.
> 
> People do that _intentionally_ with malice; they want to look like
> they know what they're talking about without having to learn what
> they're talking about.  It is intentional, self-serving debasement of
> a poor defenseless language.

  I find the words "_intentionally_ with malice" quite strong.  I'm
quite sure that there is such people, but the things is that most of the
clusters of dedicated computers made with cheap hardware for intensive
calculations that I saw was called Beowulf by their creators.  And the
name 'Beowulf' is a lot less typing for talking about this specific kind
of cluster.  I would say that most people use the term Beowulf just
because most people around them use that word and not with any malice in
mind, even when they misuse the term.

  I also found that definition from http://www.beowulf.org/ after
cliking on Beowulf FAQ:

> 1. What's a Beowulf? [1999-05-13]
> 
> It's a kind of high-performance massively parallel computer built
> primarily out of commodity hardware components, running a free-software
> operating system like Linux or FreeBSD, interconnected by a private
> high-speed network.  It consists of a cluster of PCs or workstations
> dedicated to running high-performance computing tasks.  The nodes in
> the cluster don't sit on people's desks; they are dedicated to running
> cluster jobs.  It is usually connected to the outside world through
> only a single node.
> 
> Some Linux clusters are built for reliability instead of speed.  These
> are not Beowulfs.

http://www.scyld.com/FAQs.html

>What defines a Beowulf cluster?
>
> -Off the shelf computer systems running an open source operating
> system (typically Linux) networked with a private system area network to
> function as a parallel processing super-computer
>
>What is a Beowulf cluster useful for?
>
> -Computationally intensive functions
> -Large scale processing of data
> -Manipulation of large amounts of data
> -Provides a significant improvement in the price/performance ratio as
>  compared to traditional supercomputers (up to two orders of magnitude)

  If I believe theses definitions, what i want to do _is_ actually a
Beowulf.  But you can also complain directly to Scyld Computing
Corporation for all the confusion they created if you don't agree.  I
can also talk about Beowulf-class or Beowulf-like cluster if it's all
you need to be happy.  Still, maybe many people did'nt understood that I
was _really_ meaning Beowulf (and not any kind of cluster) when I used
the word, and I am sorry if I confused anyone.


  For the peoples that questionned if it was feasable and practical to
run my simulator in a Beowulf-like cluster, I say that my problem is
very possible to parallelize (and maybe even easy for an expert).  The
calculations are mainly integration of a huge amount of differential
equations for a small delta t, but there is basically about only 100
different equations which are reused all the time the same way with
different data.  And the data are very well compartmented by groups that
occupy about only 1 KB.  Each group of data have to know the data in
only 6 others groups of data (it works in 3D), so I just have to
transfert on the network the data which are shared by more than one node
(about 60MB with 8 symetrical nodes and a simulation that takes about 1
GB of data, whitout counting all the buffers).


Simon Valiquette


"...I'm not one of those who think Bill Gates is the devil.  I simply
suspect that if Microsoft ever met up with the devil, it wouldn't need
an interpreter."
  - N. Petreley