Automatic MPI to AMPI Conversion using Photran Stas Negara, Kuo-Chuan Pan, Gengbin Zheng, Natasha Negara, Ralph Johnson, Laxmikant Kalé, Paul Ricker 8 th Annual Workshop on Charm++ and its Applications April 28, 2010
Presentation outline
Jan 20, 2016
Automatic MPI to AMPI Conversion using Photran Stas Negara , Kuo-Chuan Pan, Gengbin Zheng, Natasha Negara, Ralph Johnson, Laxmikant Kal é, Paul Ricker 8 th Annual Workshop on Charm++ and its Applications April 28, 2010. Presentation outline. MPI to AMPI code transformation Tool Evaluation - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Automatic MPI to AMPI Conversion using Photran
Stas Negara, Kuo-Chuan Pan, Gengbin Zheng, Natasha Negara, Ralph Johnson, Laxmikant Kalé,
Paul Ricker
8th Annual Workshop on Charm++ and its Applications
April 28, 2010
2
Presentation outline
• MPI to AMPI code transformation
• Tool
• Evaluation
• Future work
3
MPI to AMPI code transformation (1 of 5)
• Remove global variables
• Write pack/unpack subroutine
• Rename main PROGRAM to MPI_MAIN subroutine
4
MPI to AMPI code transformation (2 of 5)
MODULE MyMod
REAL :: p, r
INTEGER,
PRIVATE :: i, j
...
END MODULE
Fortran global variables:
• Module variables
• Saved subprogram variables
• Common block variables
SUBROUTINE MySub
REAL, SAVE :: p, r
INTEGER :: c = 0, i
...END SUBROUTINE
PROGRAM MyProg INTEGER :: i COMMON /CB/ i i = 3 CALL PrintValEND PROGRAM
SUBROUTINE PrintVal INTEGER :: j COMMON /CB/ j print *, “j=“, jEND SUBROUTINE
5
MPI to AMPI code transformation (3 of 5)
Global variables privatization:1. Generate stubs for the derived type and its
containing module2. Add an extra parameter to each
subprogram* and every call site3. Remove every global variable:
a. Declare the corresponding field in the derived type
b. Replace every access to the variable with the access to the corresponding field
c. Delete global variable
6
MPI to AMPI code transformation (4 of 5)PROGRAM MyProg INTEGER :: i COMMON /CB/ i i = 3 CALL PrintValEND PROGRAM
SUBROUTINE PrintVal INTEGER :: j COMMON /CB/ j print *, “j=“, jEND SUBROUTINE
MODULE GenMod
TYPE GenType
END TYPE
END MODULE
PROGRAM MyProg USE GenMod TYPE(GenType) :: p INTEGER :: i COMMON /CB/ i i = 3 CALL PrintVal(p)END PROGRAM
SUBROUTINE PrintVal(p) USE GenMod TYPE(GenType) :: p INTEGER :: j COMMON /CB/ j print *, “j=“, jEND SUBROUTINE
MODULE GenMod
TYPE GenType
END TYPE
END MODULE
SUBROUTINE PrintVal(p) USE GenMod TYPE(GenType) :: p INTEGER :: j COMMON /CB/ j print *, “j=“, jEND SUBROUTINE
MODULE GenMod
TYPE GenType
INTEGER :: f
END TYPE
END MODULE
PROGRAM MyProg USE GenMod TYPE(GenType) :: p INTEGER :: i COMMON /CB/ i i = 3 CALL PrintVal(p)END PROGRAM
PROGRAM MyProg USE GenMod TYPE(GenType) :: p INTEGER :: i COMMON /CB/ i p%f = 3 CALL PrintVal(p)END PROGRAM
SUBROUTINE PrintVal(p) USE GenMod TYPE(GenType) :: p INTEGER :: j COMMON /CB/ j print *, “j=“, p%fEND SUBROUTINE
MODULE GenMod
TYPE GenType
INTEGER :: f
END TYPE
END MODULE
PROGRAM MyProg USE GenMod TYPE(GenType) :: p p%f = 3 CALL PrintVal(p)END PROGRAM
SUBROUTINE PrintVal(p) USE GenMod TYPE(GenType) :: p print *, “j=“, p%fEND SUBROUTINE
MODULE GenMod
TYPE GenType
INTEGER :: f
END TYPE
END MODULE
7
MODULE GenMod
TYPE GenType
REAL, ALLOCATABLE :: ar(:)
END TYPE
END MODULE
MPI to AMPI code transformation (5 of 5)SUBROUTINE GenPUP(p, g) USE pupmod USE GenMod TYPE(GenType) :: g LOGICAL :: isAllocated INTEGER :: p, n(2) IF (.not. fpup_isunpacking(p)) THEN isAllocated = allocated(g%ar) ENDIF
CALL fpup_logical(p, isAllocated)
IF (isAllocated) THEN IF (fpup_isunpacking(p)) THEN CALL fpup_ints(p, n, 2) ALLOCATE(g%ar(n(1):n(2))) ELSE n(1) = LBOUND(g%ar, DIM=1) n(2) = UBOUND(g%ar, DIM=1) CALL fpup_ints(p, n, 2) ENDIF CALL fpup_doubles(p, g%ar, SIZE(g%ar)) IF (fpup_isdeleting(p)) THEN DEALLOCATE(g%ar) ENDIF ENDIFEND SUBROUTINE
8
Tool (1 of 2)
• Implemented in Java
• Based on Photran IDE
• Operates on Fortran 90
• Requires “pure” Fortran code
• Completely automates the transformation, except packing/unpacking of derived types
• Accessible as a refactoring in Photran
9
10
Evaluation (1 of 4)
• Evaluated on FLASH project
• Transformed 2D simulation of Sedov-Taylor explosion
• Ran experiments on NCSA Abe using 16 physical processors
• Employed GreedyLB and RefineLB
• Achieved up to 8% improvement due to load balancing
11
Evaluation (2 of 4)
12
Evaluation (3 of 4)
13
Evaluation (4 of 4)
14
Future work
• Automatically generate pack/unpack code for derived types
• Minimize overhead of the transformation
• Continue evaluation:
– More complex and larger problems
– More sophisticated load balancers
Related Documents
