6.5 Optimization of the performance

   
6.5.1 Communications

To write post-processed fields in an output file, you first gather the distributed pieces of these fields from the different processors.
Rather than gathering the fields one after the other, the fields are grouped in chunks, and these chunks are treated one after the other.
  The variable NFPXFLD is the maximum size of these chunks. Lowering it should save memory to the detriment of inter-processors communications, and vice versa.

   
6.5.2 Segmentation

Several variables are control the segmentation of the software arrays :

 

NPROMA

is the elementary size of the gridpoint rows in the model geometry. In the post-processing it is in use mostly during the vertical interpolations.  

NFPROMAG

is the elementary size of the gridpoint rows in the post-processing geometry. It is used mostly during the horizontal interpolations.  

NFPROMEL

is the elementary size of the gridpoint rows in the post-processed extension zone for LAM output. It is used only in ALADIN during the computation of the post-processed extension zone.

By definition all these variables control a part of the vectorization depth as well as memory cost. The bigger these variables are, the deeper the vectorization is, in detriment to the memory cost. On non-vector machine it is better to use small values for these parameters in order to fit the cache memory. They should not be a power of 2 to avoid memory bank conflicts. One should refer to the machine constructor to choose the best values for these variables.