Writing Scattered Data

• Problem: dump distributed particle data to disk as fast as possible
  particle data f.i. position, mass, id

  • described by structure Particle
  eType = MPI datatype corresponding to Particle
  filetype = etype
  all tasks use the same (shared) file pointer
• new MPI routines:
  MPI_File_write_shared(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status)

  • can be used seperately by every task
  • order of data depends on runtime behaviour
  MPI_File_write_ordered(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status)

  • collective operation
  • ordered by ascending rank
  both functions are not implemented in HP MPI 1.4!
• Remarks to the programs (shared1.c, shared2.c):
  two versions:

  • shared1:
    • no explicit data ordering
    • fast dump, but restore slow
    • master reads all data and distributes to tasks
  • shared2:
    • fixed order
    • restore: use cyclic distribution
  init: all tasks create known set of test data
  check: task 0 reads file in again und checks data
  MPI_Type_struct replaced by MPI_Type_create_struct

  • in C identical, in Fortran special integer type for addresses
  • not implemented in HP MPI 1.4
  possible padding in Particle prevents portability

  • file not portable as well
  • remedy: make holes explicit, use large enough holes in pType for every architecture