itaps-parallel [Fwd: some results indicating issues on interprocessor communications.]

[Mark Shephard]

We have been discussing various options for controlling communications
during operations that deal with meshes and I have been pretty load
about needing to support specific things and not do things with just one
way. To help explain why I have been making such a deal about Alex ran
some simple tests on a Blue Gene/L (slow processors, fast interprocess
communications) to look at communications costs. I am including some of
that information in my SIAM talk for which a pdf of the working draft is
attached. This study in presented in slides 22-31. You will note on
slide 26 that the ratio of communications time to total time is quite
low for refinement where we do no migration and just a minimum set of
communications. Now compare this to the more general case that includes
coarsening, where migration and more communication is required (slide
31) where communication is approaching 90% of the total time.

Mark

-------------- next part --------------
A non-text attachment was scrubbed...
Name: SIAM-Parallel-08.pdf
Type: application/pdf
Size: 3811653 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/itaps-parallel/attachments/20080309/a7ffe054/attachment.pdf>