netcdf-c/libdispatch/dparallel.c

/* Copyright 2018 University Corporation for Atmospheric
   Research/Unidata. See COPYRIGHT file for more info. */
/**
 * @file
 * @internal This file has the parallel I/O functions which correspond
 * to the serial I/O functions.
*/
#include "config.h"
#include "ncdispatch.h"
#ifdef HAVE_STDIO_H
#include <stdio.h>
#endif

/**
Create a netCDF file for parallel I/O.

This function creates a new netCDF file for parallel I/O access.

Parallel I/O access is only available in library build which support
parallel I/O. To support parallel I/O, netCDF must be built with
netCDF-4 enabled (configure options --enable-netcdf-4 and --enable-parallel4)
and with a HDF5 library that supports parallel I/O, or with support for the
PnetCDF library via the --enable-pnetcdf option.
This function is collective, i.e. must be called by all MPI processes defined
in the MPI communicator, argument comm. In addition,
values of arguments of this function must be the same among all MPI processes.

See nc_create() for a fuller discussion of file creation.

\note When opening a netCDF-4 file HDF5 error reporting is turned off,
if it is on. This doesn't stop the HDF5 error stack from recording the
errors, it simply stops their display to the user through stderr.

\param path The file name of the new netCDF dataset.

\param cmode The creation mode flag. The following flags are available:
  NC_CLOBBER (overwrite existing file),
  NC_NOCLOBBER (do not overwrite existing file),
  NC_NETCDF4 (create netCDF-4/HDF5 file),
  NC_CLASSIC_MODEL (enforce netCDF classic mode on netCDF-4/HDF5 files),
  NC_64BIT_OFFSET (create CDF-2 file),
  NC_64BIT_DATA (create CDF-5 file).

\param comm the MPI communicator specifying the processes participating the
parallel I/O to this file.

\param info MPI info object containing I/O hints or MPI_INFO_NULL.

\param ncidp Pointer to location where returned netCDF ID is to be
stored.

\returns ::NC_NOERR No error.
\returns ::NC_ENOPAR Library was not built with parallel I/O features.
\returns ::NC_EPERM: Attempting to create a netCDF file in a directory where you do not have permission to create files.
\returns ::NC_ENOTBUILT Library was not built with NETCDF4 or PnetCDF.
\returns ::NC_EEXIST Specifying a file name of a file that exists and also specifying NC_NOCLOBBER.
\returns ::NC_EINVAL Invalid input parameters.
\returns ::NC_ENOMEM System out of memory.
\returns ::NC_EHDFERR HDF5 error (netCDF-4 files only).
\returns ::NC_EFILEMETA Error writing netCDF-4 file-level metadata in
HDF5 file. (netCDF-4 files only).

<h1>Example</h1>

In this example from nc_test4/tst_parallel.c, a file is created for
parallel I/O.

\code
    int mpi_size, mpi_rank;
    MPI_Comm comm = MPI_COMM_WORLD;
    MPI_Info info = MPI_INFO_NULL;
    int ncid, v1id, dimids[NDIMS];
    char file_name[NC_MAX_NAME + 1];

    MPI_Init(&argc,&argv);
    MPI_Comm_size(MPI_COMM_WORLD, &mpi_size);
    MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);

    snprintf(file_name, sizeof(file_name), "%s/%s", TEMP_LARGE, FILE);
    if ((res = nc_create_par(file_name, NC_NETCDF4, comm, info, &ncid))) ERR;

    if (nc_def_dim(ncid, "d1", DIMSIZE, dimids)) ERR;
    if (nc_def_dim(ncid, "d2", DIMSIZE, &dimids[1])) ERR;
    if (nc_def_dim(ncid, "d3", NUM_SLABS, &dimids[2])) ERR;

    if ((res = nc_def_var(ncid, "v1", NC_INT, NDIMS, dimids, &v1id))) ERR;

        ... use collective I/O
    err = nc_var_par_access(ncid, v1id, NC_COLLECTIVE); ERR

    if ((res = nc_enddef(ncid))) ERR;

    ...

    if ((res = nc_close(ncid))) ERR;
\endcode
\author Ed Hartnett, Dennis Heimbigner
\ingroup datasets
*/
int nc_create_par(const char *path, int cmode, MPI_Comm comm,
                  MPI_Info info, int *ncidp)
{
#ifndef USE_PARALLEL
    NC_UNUSED(path);
    NC_UNUSED(cmode);
    NC_UNUSED(comm);
    NC_UNUSED(info);
    NC_UNUSED(ncidp);
    return NC_ENOPAR;
#else
    NC_MPI_INFO data;

#ifndef USE_PNETCDF
    /* PnetCDF is disabled but user wants to create classic file in parallel */
    if (!(cmode & NC_NETCDF4))
        return NC_ENOTBUILT;
#endif

#ifndef USE_NETCDF4
   /* NETCDF4 is disabled but user wants to create NETCDF4 file in parallel */
    if (cmode & NC_NETCDF4)
        return NC_ENOTBUILT;
#endif

    /* Can't use both parallel and diskless|inmemory|mmap. */
    if (cmode & (NC_DISKLESS|NC_INMEMORY|NC_MMAP))
        return NC_EINVAL;

    data.comm = comm;
    data.info = info;
    return NC_create(path, cmode, 0, 0, NULL, 1, &data, ncidp);
#endif /* USE_PARALLEL */
}

/**
Open an existing netCDF file for parallel I/O.

This function opens an existing netCDF dataset for parallel I/O access. It
determines the underlying file format automatically. Use the same call to open
a netCDF classic, 64-bit offset, CDF-5, or netCDF-4 file.

Parallel I/O access is only available in library build which support
parallel I/O. To support parallel I/O, netCDF must be built with
netCDF-4 enabled (configure options --enable-netcdf-4 and --enable-parallel4)
and with a HDF5 library that supports parallel I/O, or with support for the
PnetCDF library via the --enable-pnetcdf option.
This function is collective, i.e. must be called by all MPI processes defined
in the MPI communicator, argument comm. In addition,
values of arguments of this function must be the same among all MPI processes.

It is not necessary to pass any information about the format of the
file being opened. The file type will be detected automatically by the
netCDF library.

As of version 4.3.1.2, multiple calls to nc_open_par() with the same
path will return the same ncid value.

\note When opening a netCDF-4 file HDF5 error reporting is turned off,
if it is on. This doesn't stop the HDF5 error stack from recording the
errors, it simply stops their display to the user through stderr.

\param path File name for netCDF dataset to be opened.

\param omode The open mode flag may be NC_WRITE (for read/write
access) or NC_NOWRITE (for read-only access).

\param comm the MPI communicator specifying the processes participating the
parallel I/O to this file.

\param info MPI info object containing I/O hints or MPI_INFO_NULL.

\param ncidp Pointer to location where returned netCDF ID is to be
stored.

nc_open_par() returns the value NC_NOERR if no errors
occurred. Otherwise, the returned status indicates an error. Possible
causes of errors include:

\returns ::NC_NOERR No error.
\returns ::NC_ENOPAR Library was not built with parallel I/O features.
\returns ::NC_EPERM: Attempting to open a netCDF file where you do not have permission to write the file.
\returns ::NC_ENOTBUILT Library was not built with NETCDF4 or PnetCDF.
\returns ::NC_EINVAL Invalid parameters.
\returns ::NC_ENOTNC Not a netCDF file.
\returns ::NC_ENOMEM Out of memory.
\returns ::NC_EHDFERR HDF5 error. (NetCDF-4 files only.)
\returns ::NC_EDIMMETA Error in netCDF-4 dimension metadata (NetCDF-4 files only.)

<h1>Examples</h1>

Here is an example using nc_open_par() from examples/C/parallel_vara.c.

\code
#include <mpi.h>
#include <netcdf.h>
#include <netcdf_par.h>
   ...
    int ncid, omode;
    char filename[128];
   ...
    omode = NC_NOWRITE;
    err = nc_open_par(filename, omode, MPI_COMM_WORLD, MPI_INFO_NULL, &ncid); FATAL_ERR
\endcode
\author Ed Hartnett, Dennis Heimbigner
\ingroup datasets
*/
int
nc_open_par(const char *path, int omode, MPI_Comm comm,
            MPI_Info info, int *ncidp)
{
#ifndef USE_PARALLEL
    NC_UNUSED(path);
    NC_UNUSED(omode);
    NC_UNUSED(comm);
    NC_UNUSED(info);
    NC_UNUSED(ncidp);
    return NC_ENOPAR;
#else
    NC_MPI_INFO mpi_data;

    mpi_data.comm = comm;
    mpi_data.info = info;

    return NC_open(path, omode, 0, NULL, 1, &mpi_data, ncidp);
#endif /* USE_PARALLEL */
}

/**
This is the same as nc_open_par(), but accepts the MPI comm/info as
integers.

\param path File name for netCDF dataset to be opened.

\param omode The open mode flag may be NC_WRITE (for read/write
access) or NC_NOWRITE (for read-only access).

\param comm the MPI communicator specifying the processes participating the
parallel I/O to this file.

\param info MPI info object containing I/O hints or MPI_INFO_NULL.

\param ncidp Pointer to location where returned netCDF ID is to be
stored.

nc_open_par() returns the value NC_NOERR if no errors occurred. Otherwise, the
returned status indicates an error. Possible causes of errors include:

\returns ::NC_NOERR No error.
\returns ::NC_EINVAL Invalid parameters.
\returns ::NC_ENOTNC Not a netCDF file.
\returns ::NC_ENOMEM Out of memory.
\returns ::NC_EHDFERR HDF5 error. (NetCDF-4 files only.)
\returns ::NC_EDIMMETA Error in netCDF-4 dimension metadata. (NetCDF-4
files only.)

\author Ed Hartnett
\ingroup datasets
\internal
*/
int
nc_open_par_fortran(const char *path, int omode, int comm,
                    int info, int *ncidp)
{
#ifndef USE_PARALLEL
    NC_UNUSED(path);
    NC_UNUSED(omode);
    NC_UNUSED(comm);
    NC_UNUSED(info);
    NC_UNUSED(ncidp);
    return NC_ENOPAR;
#else
    MPI_Comm comm_c;
    MPI_Info info_c;

    /* Convert fortran comm and info to C comm and info, if there is a
     * function to do so. Otherwise just pass them. */
#ifdef HAVE_MPI_COMM_F2C
    comm_c = MPI_Comm_f2c(comm);
#else
    comm_c = (MPI_Comm)comm;
#endif
#ifdef HAVE_MPI_INFO_F2C
    info_c = MPI_Info_f2c(info);
#else
    info_c = (MPI_Info)info;
#endif

    return nc_open_par(path, omode, comm_c, info_c, ncidp);
#endif
}

/**
   @ingroup datasets

   This function will change the parallel access of a variable from
   independent to collective and vice versa.

   This function is collective, i.e. must be called by all MPI
   processes defined in the MPI communicator used in nc_create_par()
   or nc_open_par(). In addition, values of arguments of this function
   must be the same among all MPI processes.

   To obtain a good I/O performance, users are recommended to use
   collective mode. In addition, switching between collective and
   independent I/O mode can be expensive.

   In netcdf-c-4.7.4 or later, using hdf5-1.10.2 or later, the zlib,
   szip, fletcher32, and other filters may be used when writing data
   with parallel I/O. The use of these filters require collective
   access. Turning on the zlib (deflate) or fletcher32 filter for a
   variable will automatically set its access to collective if the
   file has been opened for parallel I/O. Attempts to set access to
   independent will return ::NC_EINVAL.

   @note When the library is build with --enable-pnetcdf, and when
   file is opened/created to use PnetCDF library to perform parallel
   I/O underneath, argument varid is ignored and the mode changed by
   this function applies to all variables. This is because PnetCDF
   does not support access mode change for individual variables. In
   this case, users may use NC_GLOBAL in varid argument for better
   program readability.

   @param ncid NetCDF or group ID, from a previous call to
   nc_open_par(), nc_create_par(), nc_def_grp(), or associated inquiry
   functions such as nc_inq_ncid().

   @param varid Variable ID

   @param par_access NC_COLLECTIVE or NC_INDEPENDENT.

   @return ::NC_NOERR No error.
   @return ::NC_EBADID Invalid ncid passed.
   @return ::NC_ENOTVAR Invalid varid passed.
   @return ::NC_ENOPAR File was not opened with nc_open_par/nc_create_par.
   @return ::NC_EINVAL Invalid par_access specified, or attempt to set
   filtered variable to independent access.

   <h1>Example</h1>

   Here is an example from examples/C/parallel_vara.c which changes
   the parallel access of a variable and then writes to it.

@code
#define NY 10
#define NX 4

    #include <mpi.h>
    #include <netcdf.h>
    #include <netcdf_par.h>
        ...
    char filename[128];
    int err, ncid, cmode, varid, dimid, dimid[2], buf[NY][NX];
    size_t global_ny, global_nx, start[2], count[2];
        ...
    global_ny = NY;
    global_nx = NX * nprocs;
        ...
    cmode = NC_CLOBBER;
    err = nc_create_par(filename, cmode, MPI_COMM_WORLD, MPI_INFO_NULL, &ncid); FATAL_ERR
        ...
    err = nc_def_dim(ncid, "Y", global_ny, &dimid[0]); ERR
    err = nc_def_dim(ncid, "X", global_nx, &dimid[1]); ERR
    err = nc_def_var(ncid, "var", NC_INT, 2, dimid, &varid); ERR
        ...
        ... set collective I/O globally (for all variables)
    err = nc_var_par_access(ncid, NC_GLOBAL, NC_COLLECTIVE); ERR

    err = nc_enddef(ncid); ERR
        ...
    start[0] = 0;
    start[1] = NX * rank;
    count[0] = NY;
    count[1] = NX;

    err = nc_put_vara_int(ncid, varid, start, count, &buf[0][0]); ERR

@endcode
@author Ed Hartnett, Dennis Heimbigner
*/
int
nc_var_par_access(int ncid, int varid, int par_access)
{
#ifndef USE_PARALLEL
    NC_UNUSED(ncid);
    NC_UNUSED(varid);
    NC_UNUSED(par_access);
    return NC_ENOPAR;
#else
    int stat = NC_NOERR;
    NC* ncp;

    if ((stat = NC_check_id(ncid, &ncp)))
       return stat;

    return ncp->dispatch->var_par_access(ncid,varid,par_access);
#endif
}

/**
Create a netCDF file for parallel access from the Fortran API.

This function calls nc_create_par() after converting the MPI comm and
info from Fortran to C, if necessary.

\param path The file name of the new netCDF dataset.

\param cmode The creation mode flag. The following flags are available:
  NC_CLOBBER (overwrite existing file),
  NC_NOCLOBBER (do not overwrite existing file),
  NC_NETCDF4 (create netCDF-4/HDF5 file),
  NC_CLASSIC_MODEL (enforce netCDF classic mode on netCDF-4/HDF5 files),
  NC_64BIT_OFFSET (create CDF-2 file)
  NC_64BIT_DATA (create CDF-5 file)

\param comm the MPI communicator specifying the processes participating the
parallel I/O to this file.

\param info MPI info object containing I/O hints or MPI_INFO_NULL.

\param ncidp Pointer to location where returned netCDF ID is to be
stored.

\returns ::NC_NOERR No error.
\returns ::NC_ENOPAR Library was not built with parallel I/O features.
\returns ::NC_EPERM: Attempting to create a netCDF file in a directory where you do not have permission to create files.
\returns ::NC_EINVAL Invalid input parameters.
\returns ::NC_ENOMEM System out of memory.
\returns ::NC_ENOTNC Binary format could not be determined.
\returns ::NC_EHDFERR HDF5 error (netCDF-4 files only).
\returns ::NC_EFILEMETA Error writing netCDF-4 file-level metadata in
HDF5 file. (netCDF-4 files only).

\author Ed Hartnett, Dennis Heimbigner
\ingroup datasets
\internal
*/
int
nc_create_par_fortran(const char *path, int cmode, int comm,
                      int info, int *ncidp)
{
#ifndef USE_PARALLEL
    NC_UNUSED(path);
    NC_UNUSED(cmode);
    NC_UNUSED(comm);
    NC_UNUSED(info);
    NC_UNUSED(ncidp);
    return NC_ENOPAR;
#else
    MPI_Comm comm_c;
    MPI_Info info_c;

    /* Convert fortran comm and info to C comm and info, if there is a
     * function to do so. Otherwise just pass them. */
#ifdef HAVE_MPI_COMM_F2C
    comm_c = MPI_Comm_f2c(comm);
#else
    comm_c = (MPI_Comm)comm;
#endif
#ifdef HAVE_MPI_INFO_F2C
    info_c = MPI_Info_f2c(info);
#else
    info_c = (MPI_Info)info;
#endif

    return nc_create_par(path, cmode, comm_c, info_c, ncidp);
#endif
}