Skip to content

IO and Communications

Alexandra Bozec edited this page Apr 29, 2019 · 1 revision

mod_za.F90

The following are one line descriptions of all public HYCOM I/O routines:

  zagetc  - get (read) one line from a text file

  zaiost  - initialize all array i/o.
  zaiopn  - open a file for array i/o, filename from array i/o unit.
  zaiope  - open a file for array i/o, filename from environment variable.
  zaiopf  - open a file for array i/o, filename provided.
  zaiopi  - is an array i/o unit open?
  zaiocl  - close  an array i/o unit.
  zaiofl  - flush  an array i/o unit.
  zaioiq  - query  an array i/o unit.
  zaiorw  - rewind an array i/o unit.
  zaiord3 - read   an array (3-D).
  zaiord  - read   an array.
  zaiordg - read   an array, return non-tiled (global) real*4 array.
  zaiosk  - skip   an array read.
  zaiowr3 - write  an array (3-D).
  zaiowr  - write  an array.

All these routines are assumed to be called with identical argument lists by all processors when using SPMD message passing. This is not difficult to arrange, since by default all routines are called in this manner in a SPMD run. The zaio routines are used to process HYCOM ".a" files, which contain array data only (see below). The zagetc routine is used to process HYCOM ".b" plain text files. These are only opened on the first processor, so under MPI zagetc reads a line on the first MPI task and then broadcasts it to all other tasks.

Three versions of each subroutine are provided, mod_za_mp.h for message passing with I/O from the first MPI task in each row, mod_za_mp1.h for message passing with I/O from the first MPI task, and mod_za_sm.h for a single processor (and OpenMP). The appropriate version is included in mod_za.F90 under control of cpp macros. For MPI the default is "mp", set the SERIAL_IO macro for "mp1". The routines are configured as a module, and all HYCOM routines should start with use mod_za to allow HYCOM communication routines to be invoked when required.

A special version of mod_za.F is also used in HYCOM-tools (single processor) and HYCOM-tools/MPI (one MPI task per row).

A related API with "za" replaced by "zb" is in HYCOM-tools/*/src/mod_zb.F. It is only used when reading in the full domain (via zaiord), but writing out a sub-domain (via zbiowr).

HYCOM ".a" files contain idm*jdm 32-bit IEEE real values for each array, in standard fortran element order, followed by padding to a multiple of 4096 32-bit words (16K bytes), but otherwise with no control bytes/words, and input values of 2.0**100 indicating a data void. Each record is padded to 16K bytes to potentially improve I/O performance on some machines by aligning record boundaries on disk block boundaries.

The associated HYCOM ".b" files are plain text and typically contain a 5-line header followed by one line for each 2-D array in the ".a" file. The format of the per-array line varies, but it typically use "=" to separate an array description from trailing numeric values. The latter always include the minimum and maximum value of the array as returned by zaiowr or zaiowr3 (which should be checked against the min/max of the actual array on input, returned by zaiord or zaiord3, to confirm that the .a and .b files are in fact associated with each other).


mod_xc.F90

The following are one line descriptions of all public HYCOM communication routines:

  xcaget - convert an entire 2-D array from tiled to non-tiled layout.   
  xcaput - convert an entire 2-D array from non-tiled to tiled layout.   
  xcastr - broadcast a 1-D array to all tiles.   
  xceget - find the value of a(ia,ja) on the non-tiled 2-D grid.   
  xceput - fill a single element in the non-tiled 2-D grid.   
  xciget - find the value of a(ia(:),ja(:)) on the non-tiled 2-D grid.   
  xciput - fill a list of elements in the non-tiled 2-D grid.   
  xchalt - emergency stop all processes, called by one process.   
  xclget - extract a line of elements from the non-tiled 2-D grid.   
  xclput - fill a line of elements in the non-tiled 2-D grid.   
  xcmaxr - replace array a with its element-wise maximum over all tiles.   
  xcminr - replace array a with its element-wise minimum over all tiles.   
  xcpipe - copy aa_slave to aa_master (pipe_compare with OCEANS2 only).   
  xcspmd - initialize processor data structures, called once.   
  xcstop - stop all processes, called by all processes.   
  xcsum  - sum     of a 2-D array.   
  xcsumj - row-sum of a 2-D array.    
  xcsync - barrier, no processor exits until all arrive (flush stdout).   
  xctbar - sync with processors ipe1 and ipe2 (internal use only).   
  xctilr - update the tile overlap halo of a 3-D real array.   

  xctmri - initialize timers (called in xcspmd).   
  xctmr0 - start timer.   
  xctmr1 - add time since call to xctim0 to timer.   
  xctmrn - register name of timer.   
  xctmrp - print all active timers.   
  xctmrs - set which events to time (default 1st 5000 then every 50th).   

With the exception of xchalt, All these routines are assumed to be called with identical argument lists by all processors when using SPMD message passing. This is not difficult to arrange, since by default all routines are called in this manner in a SPMD run. Most communication routines act as implicit barriers that syncronize processor state, i.e. when a processor exits a communication routine at the very least all processors that must communicate with it have entered the same subroutine. In addition the subroutine xcsync provided for cases where all processors must enter a critical section of code before the first processor exits.

Two versions of each subroutine are provided, mod_xc_mp.h for message passing, and mod_xc_sm.h for a single processor. The appropriate version is included in mod_xc.F90 under control of cpp macros. The routines are configured as a module, and all HYCOM routines should start with use mod_xc to allow HYCOM communication routines to be invoked when required.

Clone this wiki locally