SPMD programs are usually written from the perspective of a single processor, yet the intended behaviour is an aggregate computation comprising many processors running the same program on local data. Combinators, suchas map, fold, scan and multibroadcast, provide a flexible way to express SPMD programs more naturally and more abstractly at the collective level. A good SPMD programming methodology begins with a specification at the collective level, where many significant transformations and optimisations can be introduced. Eventually, however, this collective level program must be transformed to the individual level in order to make it executable on an SPMD system. This paper introduces a technique needed to make the transformation possible within a formal framework: a special collective semantics for the individual level program is required in order to justify a transformation from the collective level to the individual level. The collective semantics defines the meanings of the collective communication operations, and it allows equational reasoning to be used for deriving and implementing SPMD programs.
[1]
Gudula Rünger,et al.
Abstract Parallel Machines
,
2000,
Comput. Artif. Intell..
[2]
Selim G. Akl.
Parallel computation: models and methods
,
1997
.
[3]
David W. Walker,et al.
The Design of a Standard Message Passing Interface for Distributed Memory Concurrent Computers
,
1994,
Parallel Comput..
[4]
Jack Dongarra,et al.
MPI: The Complete Reference
,
1996
.
[5]
David B. Skillicorn,et al.
Models and languages for parallel computation
,
1998,
CSUR.
[6]
Thomas Rauber,et al.
Deriving structured parallel implementations for numerical methods
,
1996,
Microprocess. Microprogramming.