Scaling of hydro codes

Greg Lindahl lindahl at keyresearch.com
Fri Apr 11 14:26:16 EDT 2003


On Thu, Apr 10, 2003 at 10:47:04AM +0200, Wolfgang Dobler wrote:

> My question is: do others find the same type of scaling for hydro codes?
> If so, how can this be understood?

CFD can vary widely. Some algorithms are cache friendly (operator
splitting, the compute part of spectral codes), some are not (3D
operators). Sometimes the data size is huge (1+ gbytes/cpu) and
sometimes it's small enough to fit in the combined L2 caches of your
cluster.

A non-cache-friendly code won't get a great speedup when you use the
2nd cpu. This is what Craig Tierney mentioned, and you can test for
this effect using a 1-cpu and 2-cpu run.

Large data sizes mean easier network scaling. You can look at that
separately by running the code at several sizes using 1 cpu per
machine. If you increase the data size as you use more cpus, this
scaling should be nearly linear.

greg


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list