Scaling of hydro codes

[Wolfgang Dobler]

We have a 3-d finite-difference hydro code and find that the time per time
step and grid point scales almost linearly,
  t_step ~ Ncpu^(-1) ,
on an Origin3000 from 1 up to 64 CPUs.

On our Linux cluster (Gbit ethernet, 8x2 CPUs) however, we get a scaling
that is well represented by
  t_step ~ Ncpu^(-0.75) .
More or less the same scaling is obtained on another machine (100Mbit, 128
nodes), and also for another hydro code (parallelized using Cactus).
Note that the number of grid points was adapted for these timings, so that
the problem size per CPU is roughly constant.

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

I don't expect latency to play a role for these timings, as we are only
communicating a reasonably low number of large arrays in every time step;
I suppose, Cactus does the same.
  And if saturation of the switch played a role, I would expect a
well-defined drop at some critical value of Ncpu, not a power law.


W o l f g a n g

-- 

 -------------------------------------------------------------------------
|  Wolfgang Dobler                           Phone: ++49/(0)761/3198-224  |
|  Kiepenheuer Institute for Solar Physics   Fax:   ++49/(0)761/3198-111  |
|  Schöneckstraße 6                                                       |
|  D-79104 Freiburg                   E-Mail: Dobler at kis.uni-freiburg.de  |
|  Germany                       http://www.kis.uni-freiburg.de/~dobler/  |
 -------------------------------------------------------------------------
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf