[MITgcm-support] Optimizing the 2D pressure solver

[Christopher L. Wolfe]

Michael,

The domain decomposition is dependent on the number of processors, but 
unless I'm running on a vector processor I tend to make square tiles. 
Thus, for 1024 processors, the tiles are 28 x 28 x 20 with three grid 
points for overlap in each direction. With the overlap points the tiles 
are more like 34 x 34 x 20 and roughly a third of each tile is overlap 
points.

Matt Mazloff is probably correct that ratio of overlap points to "real" 
points is getting rather large and will probably thwart attempts to 
scale to much more than 1024 processors. On the other hand, the pressure 
solver is still dominating the time-to-solution and if Ralf Giering 
managed to significantly improve its performance (as mentioned by 
Dimitris) the improvement in the GCM performance might be worth the time 
it takes to implement the changes.

Christopher

Michael Schaferkotter (Contractor) wrote:
> greetings;
>
> i looked over the preceding threads for this topic and found no mention of tile sizes or overlap size.
> aren/t  these specified in code/SIZE.h?
>
> so how is the domain being decomposed?
>
> michael
>
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support
>