Concepts

• Logical file view:
  etype:

  • basic unit of data access and file positioning
  • can be any predefined MPI datatyp
  • derived datatypes possible, if:
    • typemap displacements non-decreasing
    • no negative displacements
  filetype:

  • sequence of etypes, possibly with holes in etype size
  view:

  • defines the file data visible by a task
  • given by etype, filetype and offset to beginning of file (in bytes)
  • different views of a file possible (f.i. for different sections)
  • 
  • global distribution of data is given by combination of all task views:
  • 
• Access modes:
  positioning:

  • individual file pointer for each task (e.g. MPI_File_read)
  • one file pointer common to all tasks (e.g. MPI_File_read_shared) (requires same file view for all tasks)
  • explicit offset (e.g. MPI_File_read_at)
  synchronism:

  • blocking (e.g. MPI_File_read)
  • non-blocking (e.g. MPI_File_iread/ MPI_Wait)
  coordination:

  • individual (e.g. MPI_File_read)
  • collectiv (e.g. MPI_File_read_all)
• Physical file view:
  information about the "physical" file in MPI_Info
  some reserved keys, f.i.:

  • file_perm	permissions at file creation
    io_node_list	list of i/o-devices for the file
    striping_factor	number of devices for disk striping

• Portability:
  representation of data in a file in general architecture dependent
  because of additional format conversions, we have:

  • better portability lower performance
  parameter datarep at definition of fileview allows different portability levels:

  • native:
    • internal (memory) format
    • for a homogeneous environment
  • internal:
    • useful in heterogeneous MPI environments
  • external32:
    • architecture independent format
    • portable between different MPI implementations