Using land suppression can give different results

[charris]

At least for ORCA025, turning on land suppression gives different results on some decompositions (for example 32x30). Is this to be expected? It seems like undesirable behaviour.

[smasson]

it is not really expected... :-)
Do you still have the problem if you remove the sea-ice (usually problems comes from this tricky part...)?

[charris]

I was running without CICE or LIM (just with nn_ice=0) so it wasn't coming from the sea-ice. The differences seemed to be small but I didn't have time to investigate where they were originating.

[dupontf]

Without changing the MPI tile definition, i.e. a simple roundoff error change?

[charris]

The domain decomposition should have been the same in both cases, and I was running with key_mpp_rep so I would have expected to be able to get identical results.

[smasson]

If you compiled with ifort, did you use the option "-fp-model precise"? Without this magic option, I had different results every time I was running the model (same MPI domain decomposition)...

[charris]

I was using mpxlf90_r on our Power7. But I can normally get the same results when I change domain decomposition, so I'm fairly sure this issue is somehow related to the land suppression.

[charris]

As I say these differences are very small. Probably bit level on the first time-step (they take three or four time-steps to show up as differences in solver.stat) but I've not investigated in detail.

Has anyone else ever checked that you get the same results with land suppression?