Problems with averages_detide and OpenMp

[luzgarcia]

Hi everybody,

I have tried to run my model using OpenMp in a computer with 8 processors (domain partition NtileI=2, NtileJ=4) using the AVERAGES_DETIDE option. I checked I am using the last version of the code, which solved some bugs reported in the forum (see, for instance this post). I am using the tidal forcing in an independent netcdf file with tide_period as record dimension according to this post and NRREC == -1 for the tide harmonics to be accumulated in the forcing file (I read this here). The ININAME in the ocean.in points to the previously saved restart.

What I get as output is shown in the attached Figure_openmp, that corresponds to variable zeta_detided after fifteen days of simulation (it is being saved every 24h). It seems that every tile gets its own results that have nothing to do with the rest of the tiles...

Then I decided to try to run the simulation in sequential mode making sure that NtileI=1 and NtileJ=1. Then, the result for the same date looked good (see Figure_sequential)...

And also tried with MPI and 16 processors (NtileI=4 and NtileJ=4). The results looked good as well (see Figure_mpi).

I dont know if someone has detected something like this...is it possible that there is a bug related with OpenMp or am I doing something wrong??

Thank you very much!

[arango]

Well, it seems that we have a shared-memory bug for this option. We need to invert a matrix for the detiding least-squares problem. Inverting matrices in shared-memory is not trivial and should be done by the master thread. I will add this problem to my very long to-do list. For now, just use distributed-memory for this application. Thank you for reporting this.

[linzhenhua]

Hi

From my limited knowledge, if one use domain decomposition, one have to use MPI, not OpenMP. This maybe the reason for the problem encountered above in my opinion.

Please correct me if I'm wrong.