Open aekiss opened 1 year ago
I think we should investigate this before we start doing long runs / cmip7 runs. If they are only getting calculated once after a restart and then stored for the rest of the run that might be ok, but if they are recalculated for every interpolation / regrid then we should definately pre-compute them.
Another important detail is the parallelisation of the mappings computation. Is it serial code? Parallel? If parallel, what is the scalability?
Online remapping weight generation is parallel, according to slide 5 of @mvertens CW2023 talk.
@micaeljtoliveira - that is correct. Its totally parallel.
@micaeljtoliveira - all IO in CMEPS is parallel as well.
@micaeljtoliveira do your scaling tools allow checking how well the weight generation parallelises?
do your scaling tools allow checking how well the weight generation parallelises?
Yes, we might just need to add some profiling regions.
For performance we might want to pre-compute mappings rather than auto-generating them at runtime (see https://github.com/COSIMA/MOM6-CICE6/pull/5) - but only if auto-generating consumes a lot of walltime.
Unclear whether this is currently possible with CDEPS; might require code changes.