Minor Extensions

Describes some small extensions, some of which are needed in the following sections.

• Startup:
  Many MPI implementations provide a script mpirun to start an MPI application
  recommended in MPI-2, not required:

  • mpiexec -n <numprocs> <program>
  Suggested syntax for applications using different architectures/executables:

  • via command line parameters or configuration file
  • parameters similar to MPI_Comm_spawn
• Initialization of MPI
  MPI_Init in C needs command line parameters argc, argv
  may not be available in libraries using MPI
  alternative in MPI-2:

  • argc = NULL; argv = NULL;
• System dependent information
  MPI_Info: Object for all kinds of system dependent information
  Examples:

  • file access rights for MPI_File_open
  • working directory of new tasks spawned with MPI_Comm_spawn
  contains pairs of strings (key, value)

  • MPI-2 reserves keys and their meaning for some MPI routines
  • MPI_INFO_NULL as default value, always allowed
  Routines to manage info objects:

  • int MPI_Info_create(MPI_Info *info)
  • int MPI_Info_set(MPI_Info info, char *key, char *value)
  • int MPI_Info_free(MPI_Info *info)
• Memory Allocation
  some architectures use special kinds of memory
  Examples:

  • globally shared memory
  • shared memory on one SMP node
  useful for some operations (faster memory access)
  sometimes necessary when using one-sided communication (i.e. on HP-UX)
  Allocation functions similar to malloc/free:

  • int MPI_Alloc_mem(MPI_Aint size, MPI_Info info, void *baseptr)
  • int MPI_Free_mem(void *base)
• Creation of datatypes
  Modifications and additions

  • new functions (using MPI_Aint consequently) for explicit addresses
  • simpler manipulation of upper/lower bound
  • datatypes for submatrices (useful for MPI-I/O)
  • datatypes for distributed matrices (block-cyclic distribution)
  Example: datatype for a column block of a matrix

  • 

  • MPI_Datatype columntype; 
    int sizes[2]     = {100, 100}; 
    int subsizes[2]  = {100,  25}; 
    int starts[2]    = {  0,   0}; 
    int rank; 
     
    MPI_Comm_rank(MPI_COMM_WORLD, &rank); 
    starts[1]=rank*subsizes[1]; 
     
    MPI_Type_create_subarray(2, sizes, subsizes, starts,
                       MPI_ORDER_C, MPI_DOUBLE, &columntype);

• Mixing programming languages in an MPI application
  Initialization/Termination:

  • valid independent of used language
  Transfer of handles:

  • explicit functions in C to convert from and to Fortran
  • C datatyp MPI_Fint to contain Fortran integer
  • Example Communicator:
    • MPI_Comm MPI_Comm_f2c(MPI_Fint comm)
    • MPI_Fint MPI_Comm_c2f(MPI_Comm comm)
  Message passing:

  • Types for send and receive have to be identical
    (i.e. both MPI_INT or both MPI_INTEGER)
  • portable method:
    • define special datatype for the message
    • convert handle to this datatype into other language
  • non-portable:
    • implementations may allow mixing
    • e.g. if C-int = FORTRAN-INTEGER:
      send MPI_INTEGER in Fortran, receive it as MPI_INT in C