Permanent Data Movement (original) (raw)

Up: Fortran Support Next: Comparison with C Previous: Temporary Data Movement and Temporary Memory Modification

A Fortran compiler may implement permanent data movement during the execution of a Fortran program. This would require that pointers to such data are appropriately updated. An implementation with automatic garbage collection is one use case. Such permanent data movement is in conflict with MPI in several areas:

• MPI datatype handles with absolute addresses in combination with MPI_BOTTOM.
• All nonblocking MPI operations if the internally used pointers to the buffers are not updated by the Fortran runtime, or if within an MPI process, the data movement is executed in parallel with the MPI operation.

This problem can be also solved by using the ASYNCHRONOUSattribute for such buffers. This MPI standard requires that the problems with permanent data movement do not occur by imposing suitable restrictions on the MPI library together with the compiler used; see Section Requirements on Fortran Compilers .

ExampleUsing separated variables for overlapping communication and computation to allow the protection of nonblocking communication with the ASYNCHRONOUS attribute.

USE mpi_f08
REAL :: b(0:101) ! elements 0 and 101 are halo cells REAL :: bnew(0:101) ! elements 1 and 100 are newly computed INTEGER :: i CALL separated_sections(b(0), b(1:100), b(101), bnew(0:101)) i=1 ! compute leftmost element
bnew(i) = function(b(i-1), b(i), b(i+1))
i=100 ! compute rightmost element
bnew(i) = function(b(i-1), b(i), b(i+1))
END

SUBROUTINE separated_sections(b_lefthalo, b_inner, b_righthalo, bnew)
USE mpi_f08
REAL, ASYNCHRONOUS :: b_lefthalo(0:0), b_inner(1:100), b_righthalo(101:101) REAL :: bnew(0:101) ! elements 1 and 100 are newly computed TYPE(MPI_Request) :: req(4) INTEGER :: left, right, i CALL MPI_Cart_shift(...,left,right,...) CALL MPI_Irecv(b_lefthalo ( 0), ..., left, ..., req(1), ...) CALL MPI_Irecv(b_righthalo(101), ..., right, ..., req(2), ...) ! b_lefthalo and b_righthalo is written asynchronously. ! There is no other concurrent access to b_lefthalo and b_righthalo.
CALL MPI_Isend(b_inner( 1), ..., left, ..., req(3), ...) CALL MPI_Isend(b_inner(100), ..., right, ..., req(4), ...)

DO i=2,99 ! compute only elements for which halo data is not needed bnew(i) = function(b_inner(i-1), b_inner(i), b_inner(i+1))
! b_inner is read and sent at the same time. ! This is allowed based on the rules for ASYNCHRONOUS. END DO CALL MPI_Waitall(4,req,...) END SUBROUTINE

Up: Fortran Support Next: Comparison with C Previous: Temporary Data Movement and Temporary Memory Modification

Return to MPI-3.1 Standard Index
Return to MPI Forum Home Page

(Unofficial) MPI-3.1 of June 4, 2015
HTML Generated on June 4, 2015