[Beowulf] Cluster for Finite Element Analysis

Lombard, David N david.n.lombard at intel.com
Mon Apr 25 07:10:32 PDT 2005


First question:  What code(s) will you be running: commercial (ISV) or
your own?

Second question:  What type of analysis: explicit or implicit?

For implicit ISV codes, CPU performance (especially FP) and disk I/O
bandwidth will be most significant; network I/O will not be significant.

For explicit ISV codes, the CPU is still important, then network I/O is
most significant; disk I/O will not be significant.

Memory bandwidth followed by memory size are the next most significant
in either case.

For ISV codes, look at the various benchmarks that are published BUT
make sure you look for benchmarks of the analysis types and model sizes
you intend to run.  This is very important!  Best would be the ability
to benchmark your own models, but, it can be problematic getting the
right input data from your model for the particular solver you're
looking at.

The importance of the benchmarks (or thorough understanding of the
performance characteristics of the code) cannot be understated!  For
example, the disk I/O requirement list above for implicit ISV codes can
be greatly amplified or attenuated depending on the types of analyses
and the model sizes that you run; on the other hand, the disk I/O
requirement will never likely be a significant issue for explicit codes.

Once you've looked at the types of models, and the performance of the
ISV apps, you should have a sufficient understanding to answer your
hardware questions.

If you're building you own software, then you will be the best source of
understanding the hardware requirements.

As for clustering software, there are various alternatives, commercial,
e.g., Scyld; opens source cluster stacks, e.g., OSCAR, NPACI Rocks,
Clustermatic, Warewulf; or open-source roll-your-own.  Do a little
surfing to see what suits your needs *and* you application.  Having said
that, unless you're fundamentally interested in a learning exercise, and
have the time, I'd not yet roll your own, go for an existing cluster
stack.

Finally, as for your last question, the cluster stack will have a
cluster build-out process, use that.

But, the number one rule is: Choose the hardware based on the needs of
the application!  If you build the hardware first, and then choose the
application, well, you may be lucky and happy, but you will more likely
wish you had spent your hardware dollars differently.

-- 
David N. Lombard
 
My comments represent my opinions, not those of Intel Corporation.




More information about the Beowulf mailing list