PARALLEL.RU - Информационно-аналитический центр по параллельным вычислениям

9.5.1. Datatypes for File Interoperability

If the file data representation is other than ``native,'' care must be taken in constructing etypes and filetypes. Any of the datatype constructor functions may be used; however, for those functions that accept displacements in bytes, the displacements must be specified in terms of their values in the file for the file data representation being used. MPI will interpret these byte displacements as is; no scaling will be done. The function MPI_FILE_GET_TYPE_EXTENT can be used to calculate the extents of datatypes in the file. For etypes and filetypes that are portable datatypes (see Section Semantic Terms ), MPI will scale any displacements in the datatypes to match the file data representation. Datatypes passed as arguments to read/write routines specify the data layout in memory; therefore, they must always be constructed using displacements corresponding to displacements in memory.

[] Advice to users.

One can logically think of the file as if it were stored in the memory of a file server. The etype and filetype are interpreted as if they were defined at this file server, by the same sequence of calls used to define them at the calling process. If the data representation is ``native'', then this logical file server runs on the same architecture as the calling process, so that these types define the same data layout on the file as they would define in the memory of the calling process. If the etype and filetype are portable datatypes, then the data layout defined in the file is the same as would be defined in the calling process memory, up to a scaling factor. The routine MPI_FILE_GET_FILE_EXTENT can be used to calculate this scaling factor. Thus, two equivalent, portable datatypes will define the same data layout in the file, even in a heterogeneous environment with ``internal'', ``external32'', or user defined data representations. Otherwise, the etype and filetype must be constructed so that their typemap and extent are the same on any architecture. This can be achieved if the they have an explicit upper bound and lower bound (defined either using MPI_LB and MPI_UB markers, or using MPI_TYPE_CREATE_RESIZED). This condition must also be fulfilled by any datatype that is used in the construction of the etype and filetype, if this datatype is replicated contiguously, either explicitly, by a call to MPI_TYPE_CONTIGUOUS, or implictly, by a blocklength argument that is greater than one. If an etype or filetype is not portable, and has a typemap or extent that is architecture dependent, then the data layout specified by it on a file is implementation dependent.

File data representations other than ``native'' may be different from corresponding data representations in memory. Therefore, for these file data representations, it is important not to use hardwired byte offsets for file positioning, including the initial displacement that specifies the view. When a portable datatype (see Section Semantic Terms ) is used in a data access operation, any holes in the datatype are scaled to match the data representation. However, note that this technique only works when all the processes that created the file view build their etypes from the same predefined datatypes. For example, if one process uses an etype built from MPI_INT and another uses an etype built from MPI_FLOAT, the resulting views may be nonportable because the relative sizes of these types may differ from one data representation to another. ( End of advice to users.)

MPI_FILE_GET_TYPE_EXTENT(fh, datatype, extent)
[ IN fh] file handle (handle)
[ IN datatype] datatype (handle)
[ OUT extent] datatype extent (integer)

int MPI_File_get_type_extent(MPI_File fh, MPI_Datatype datatype, MPI_Aint *extent)
MPI_FILE_GET_TYPE_EXTENT(FH, DATATYPE, EXTENT, IERROR)
INTEGER FH, DATATYPE, IERROR
INTEGER(KIND=MPI_ADDRESS_KIND) EXTENT
MPI::Aint MPI::File::Get_type_extent(const MPI::Datatype& datatype) const

Returns the extent of datatype in the file fh. This extent will be the same for all processes accessing the file fh. If the current view uses a user-defined data representation (see Section User-Defined Data Representations ), MPI uses the dtype_file_extent_fn callback to calculate the extent.

[] Advice to implementors.

In the case of user-defined data representations, the extent of a derived datatype can be calculated by first determining the extents of the predefined datatypes in this derived datatype using dtype_file_extent_fn (see Section User-Defined Data Representations ). ( End of advice to implementors.)