Compiling High Performance Fortran for distributed-memory architectures

The design and implementation of a parallel array operator for the arbitrary remapping of data

“…Unlike ZPL, HPF does not restrict operations between arrays with different data distributions. Thus, a statement like C = A + B may or may not require communication depending on the distributions of A, B, and C. A great deal of research in the HPF community has focused on generating efficient and minimal communication for HPF programs [13,5]. Some of these efforts apply strength reduction to communication, replacing general all-to-all communications with reductions or nearest neighbor exchanges.…”

Gilgamesh: A Multithreaded Processor-In-Memory Architecture for Petaflops Computing

“…We mention only a few among the many relevant references. In particular, languages such as Kali, Vienna Fortran, Fortran D, HPF, and Opus [26,9,14,21,10], and compilation systems such as SUPERB, the Fortran D compiler, Paradigm, and the Vienna Fortran Compiler [35,14,4,5] as well as the huge amount of work on irregular problems have strongly influenced the macroserver design. Macroservers address some weaknesses of existing systems (in particular with respect to performance), support a higher-level view of data structures and distribution than existing in most present systems, and address a number of specific features of PIM-based architectures.…”

“…The elements whose value is 0 are represented by empty boxes; the non-zero elements are explicitly specified and numbered in row major order (for simplicity, we have chosen as the value its position in this numbering scheme). The distribution is irregular, based on (2, 7), (3,8), (4, 6), (7,7); and λ(4) = [8 : 10, 5 : 8] with nonzero elements (8,5), (8,8), (9,5), (10,7).…”

“…As a consequence, the binding of collections to their distributions is a runtime action, and a representation of the distributed collection must be kept and managed at runtime. Distribution descriptors serve that purpose; they are generalizations of layout descriptors as used in HPF compilation systems (see, for example [5]); they may also contain information about access patterns (iterators) and assertions related to optimization. Finally, a distribution descriptor may be used to actually guide the dynamic generation of communication, using its encoded knowledge of the distribution and the collection and architecture index domains.…”

“…A more complete discussion of the generalized approach outlined in this paper for macroservers can be found in [38]. Work on the implementation of macroservers can draw in part on a rich body of experience with university prototypes as well as industrial software systems that were developed around the languages mentioned above, in particular implementations of HPF-like languages for distributedmemory systems (see, for example [11,16,5]). The last of these papers contains an extensive set of references to related compilation work.…”