[Beowulf] Heat transfer simulation

Mark Hahn hahn at physics.mcmaster.ca
Tue Apr 12 11:43:27 PDT 2005


> (www.umss.edu.bo) of Bolivia (south america),  my career aren't too
> apply about optimization and simulation as far i know, actually is
> there a work about "routes", finding the short path between two
> points.

I understand this.  my point is that you can't find routes 
without actually *needing* to travel them.  unless you have real
(not made up) research problems, I don't believe you can design
or build a cluster well.  you can build something very nice,
but I believe it will need to change once a real application appears.
I suggest iteration to solve this as well as most other problems.

> simulations are too slow ", but we have to. I want to be a researcher
> about modeling real problems and using computational resources.

OK, so what I would suggest doing is to run your models on 
whatever computers you have available, right now.  don't think 
too far ahead about what your future cluster will be like - 
you don't really know what it *should* be like yet.  then run 
your code in parallel on two machines with a back-to-back cable.
how does it scale?  what packet sizes does it send?  how frequently?
are you using MPI collective operations?

the details of your initial test cluster doesn't matter much - 
only that you try real codes on it, evaluate its performance, and 
learn from it.  improvement comes from iteration, not more design.


> PD: If here were researchers about high performance computing and
> applied math with computational resources, i'll not asking here.

I don't quite understand that.  I think the beowulf list is mainly
people who blur the line between builders and users of HPC clusters 
using commodity components.  there's no special affinity for or against
applied math, for instance.

regards, mark hahn.




More information about the Beowulf mailing list