[Beowulf] Simulation for clusters performance

[Dave Love]

Scott Atchley <e.scott.atchley at gmail.com> writes:

> You may also want to look at Sandia's Structural Simulation Toolkit
> <https://sst.sandia.gov> and Argonne's CODES
> <https://www.mcs.anl.gov/projects/codes/>.

And maybe smpi from SimGrid <http://simgrid.gforge.inria.fr/>, if you
can get or synthesize calibration data.  The MPI (or its tuning) may be
rather important compared with the fabric, and is rarely properly
reported in benchmark results ☹.  SimGrid, at least, can use different
collective algorithms; I don't know about the alternatives, or to what
extent these can model offloading and the various routing schemes in
use.  A comparison would be interesting if anyone has one.

<https://ieeexplore.ieee.org/document/7877012> has results for different
topologies and concludes that there's no generally best network.