[Beowulf] Re: Redmond is at it, again

[Robert G. Brown]

On Tue, 8 Jun 2004, Suvendra Nath Dutta wrote:

> How does MATLAB generate parallel code? I thought Cleve Moler 
> specifically ruled out a parallel MATLAB? I know there are efforts (at 
> least I know two from MIT & Lincoln labs) to make MATLAB run in 
> parallel, but both these run on Linux. Cornell was supposed to have one 
> version that ran on Windows, but I was told that it is now a defunct 
> project.

People (from what I've been told) use matlab to run batch jobs that are
embarrassingly parallel.  There are also libraries that one can use to
use MPI in a matlab computation, but I think that this sort of usage is
fairly primitive and probably little different from using batch jobs.

I think that people also make a sort of matlab pipeline and use it to
visualize data, to process data, to process and transform and visualize
data in a stream.

Part of the problem is that tools like matlab at their best are inferior
to writing real parallel code for long running production numerical
code.  If your research will take six months or a year running
essentially serial matlab jobs, it is probably worth your while to use
three of those months to write a C version that runs in one month or
even one week in truly parallel mode.  You can still use matlab, after
all, to do the end stage visualization if it works well there.

Scripting languages aren't toys, but neither are they ideal high
performance computing vehicles, for the most part.

   rgb

> 
> Suvendra.
> 
> 
> On Jun 4, 2004, at 1:30 AM, Douglas O'Flaherty wrote:
> 
> >
> > I wanted to add a 5th point to Laurence's excellent  summary that I 
> > didn't see elsewhere.
> >
> > (5) The developers require an environment for rapid iterations of code 
> > development targetted to clusters.
> >
> > Where Windows clusters have already been successful is in transfering 
> > PC code to a compute farm for more efficient processing. Financial 
> > Analysis using Monte Carlo comes immediately to mind. Tools such as 
> > Matlab (and in another generation MSFT Visual Studio) can create code 
> > for distributed computing. It's unlikely to be especially good MPI 
> > code, but I bet it will work. Relative to QCD or modeling, Monte Carlo 
> > is barely cluster computing, but it is a large and viable commercial 
> > market. This is a similar driver for Xserves in the data centers of 
> > Life Science researchers -- they already have Macs on their desks and 
> > they can turn the code quickly.
> >
> > The list has discussed the sysadmin and economic point of views very 
> > well, but there hasn't been much about how tools for the end-user will 
> > influences the decisions. Joe captured some of this referring to the 
> > ISV support.
> >
> > When Cornell deployed the Windows clusters back a few years, the 
> > typical cluster developer grumbled some. The surprise was how many new 
> > users asked to be on the Windows clusters. Many were not from the hard 
> > sciences, but from economics and other departments who needed to do 
> > numerical analysis. For the academics on this list, that may be an 
> > un-tapped market for funds from other departments. You might learn to 
> > like having a Windows cluster in the data center. I suggest calling it 
> > Bofa ;)
> >
> > doug
> >
> > _______________________________________________
> > Beowulf mailing list, Beowulf at beowulf.org
> > To change your subscription (digest mode or unsubscribe) visit 
> > http://www.beowulf.org/mailman/listinfo/beowulf
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
> 

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu