Minutes of the WP3a meeting (DeBilt, January 13, 2004)

(S. Valcke - 15/01/2004)

People who attended the meeting:

From NEC-CCRL, Rene Redler, Hubert Ritzdorf; from SGI Deutschland, Reiner Vogelsang; from NEC, Thomas Schoenemeyer; from CERFACS, Sophie Valcke.

• Philippe Bourcier will not be reemployed by IPSL for the PRISM last year. We are now in the process of trying to establish a contract between IPSL and CERFACS, so that Damien Declat's contract could be extended.

• Report on the development of the PSMILe, regarding the coupling part, was given by René and Hubert. Report on the development of the PSMILe regarding the I/O part was given by Reiner. Report on the development of the Driver and Transformer and SASA tools was given by Sophie.

• Sophie will establish the list of priorities for the developments to be done until the end of the project.

• Some details regarding the actual state of development:
  • Lag concept and coupling restart not included yet.

  • REAL, DOUBLE, INTEGER fields are supported in: 1- PSMILe API routines and internal structures, 2- Transformer routines, 3- I/O. All types of fields are not supported for PSMILe sending or receiving actions (direct or via Transformer); this should be added for the prototype.

  • Bundles and vectors are supported in 1- PSMILe API routines and internal structures, 2- I/O. They are not supported neither in PSMILe sending or receiving actions (direct or via Transformer), neither in the Transformer.

  • Coupling or I/O of fields after reduction is not supported and need to be specified further.

• Regarding the prototype, the following was agreed:
  • The target date for the official PRISM release is February 15th 2004.

  • Support of REAL, DOUBLE, INTEGER coupling fields need to be included in neighborhood search and PSMILe sending or receiving actions (direct or via Transformer). Hubert.

  • Interpolation of integers will be supported; whether it makes sense or not will be left to the user. (We may change this for the final version). Hubert, Damien.

  • Unstructured grids will not be supported for coupling.

  • I/O supports structured and nonstructured grids, as long as the local partitions can be expressed with only an offset in the global index space. For unstructured grids, this is the case if the local partition covers one segment in a 1D global space.

  • Single and multiple I/O are supported for inputs and outputs. Single I/O means reading from or writing to one file; multiple I/O means reading from or writing to different files (one for each process).

  • Parallel non-exact neighborhood search will be included for
    • 3D N nearest neighbour
    • trilinear
    • 2D N nearest neighbour
    • bilinear
    The 2D N nearest neighbour and bilinear will be used for grids having a vertical dimension of 1 for which the user specifies, in the SMIOC, a 2D1D interpolation with the 1D vertical interpolation being ``none'' (see SMIOC information). Hubert.

  • For trilinear interpolation, 8 neighbours will be transfered between PSMILe and Transformer. For target points near the border of a partition, a neighbour address of 0 will mean a non valid neighbour. Hubert.

  • prism_def_partition API should be redefined so to cover later unstructured grids. Sophie, Reiner.

  • The transfer of SCC information between the Driver and PSMILe should be rewritten to transfer only the information relevant for the application. Wrapping routines (analoguous to SMIOC transfer routines) should be written. Damien.

  • Data format for I/O: 3 formats (NetCDF, binary, ascii) are supported for outputs. Only NetCDF is supported for inputs.

  • PSMILe-Transformer interaction needs to be completed by using Damien's new routines: psmile_trs_finalize.F90, psmile_trs_put_int.F90, psmile_trs_get_dble.F90, psmile_trs_put_real.F90, psmile_trs_get_int.F90, psmile_trs_set_src_epio3d_dble.F90, psmile_trs_get_real.F90,

    psmile_trs_set_src_epio3d_real.F90, psmile_trs_give_neighbors3d.F90,

    psmile_trs_set_tgt_epio3d_dble.F90, psmile_trs_inform.F90,

    psmile_trs_set_tgt_epio3d_real.F90, psmile_trs_put_dble.F90,

    psmile_trs_set_triple_links.F90. Hubert.

  • Driver-PSMILe interaction for SMIOC info needs to be completed. Damien will finalize the wrapping routines, i.e PSMILe_smioc_init and PSMILe_smioc_finalize. René will include them in prism_init_comp. Damien, René.

  • Use of SMIOC information needs to be implemented in PSMILe. René.

  • Damien is currently working on the SASA routines. Depending on the difficulty of the development, they will be or not included in the prototype. Damien.

  • Only the proleptic Gregorian calendar is supported right now and will be in the prototype.

  • Calendar routines as defined on

    http://www.ccrl-nece.de/redler/PRISM/WP3A/PSMILe/ are still missing and will not be included in the prototype.

• Connectivity between grid points is needed for treatment of unstructured grids. PRISM community will be asked how connectivity is expressed in their model. ESMF method, if any, will be inquired. Sophie.

• If/when additional grid information will be required in the PSMILe/Transformer interaction (areas or angles, for example), they should all reside in one array(epio_size, nbr_info), where nbr_info is the number of additional information transfered (e.g. 2 for areas and angles).

• All grid information, except the masks, (e.g. lats, lons, areas, angles, ...) should be transfered between the PSMILe and the Transformer with the same precision (i.e. REAL or DOUBLE for all). The precision chosen should be the highest among the precisions used for the transfer of information between the model and the PSMILe. The coupling fields may have a different precision.

• For I/O, the SMIOC dtd should be revised so that a unique choice between ``single'' or ``multiple'' should be possible for elements ``fileset'' and ``threading''. Sophie.

• For I/O, the SMIOC dtd should be revised so to include an additional element indicating that the grid information should not be written to the file. Sophie.

• Time interpolation for I/O inputs:
  • If there is no user defined time interpolation, the I/O should take the date bounds into account.
    • If there is one file timestep falling in the date bounds it should be taken as input.
    • If there is no file timestep falling in the date bounds, the nearest neighbour in time should be taken as input.
    • If there is more than one timestep falling in the date bounds, a weighted average of these timesteps should be taken as input.

  • If there is a user defined time interpolation (time nearest neighbour or time linear weighted average), the input should not consider the date bounds.

• An upgrade of mpp_IO to make use of parNetCDF (parallel NetCDF based on MPI_IO) is planned. Reiner.

• Input dates and times are required in true dates and time (i.e. months $\geq$ 12, hours $\geq$ 24, minutes $\geq$ 60 and seconds $\geq$ 60 are not allowed.

• Transformer parallelisation:
  • Thomas will probably visit CERFACS beginning of February to discuss Transformer parallelisation in more details with Damien.

  • The number of processes for the Transformer would be chosen by the user.

  • The load balancing of the different Transformer processes should be automatically set at the beginning of the run by the Transformer master process. Before sending an EPIO definition, the PSMILe could send the number of points contained in the EPIO to the Transformer master process, which would then attribute the EPIO treatment to a particular Transformer process depending of the number of EPIO points already treated by the different processes.