Teaching MPI suggestion

[Greg Lindahl]

I wanted to reply on this topic, but I tossed the messages and the
beowulf.org archive of this list is 6 weeks behind, so...

If you're teaching a class about programming clusters, it might be
nice to not only cover MPI, but also some other techniques for
programming which run well on clusters but hide more of the details.

For example, the "Scalable Modeling System" from the Forecast Systems
Lab is one such system. It works well on stencil algorithms such as
finite difference weather programs: you add some HPF-style directives
to your code to distribute your data, and you explicitly mark parallel
sections and where data needs to be exchanged.

The result is highly efficient programs, and there is enough debugging
support that I think students could take an existing example serial
code and parallelize it as a class project.

The SMS homepage is at:

http://www-ad.fsl.noaa.gov/ac/sms.html

SMS has some limits. It works for a limited range of disciplines,
although more people use stencil codes than you might think. And the
only language supported is F77. That ought to be OK for a class
project: if you hand your students a working example with all the I/O
already done, they shouldn't have any problem fiddling with equations
and loops.

-- g