contributed;articles

Doi: 10.1145/2018396.2018415

HPF pioneered a high-level approach
to parallel programming but failed
to win over a broad user community.

By keN keNNeDy, chaRLes koeLBeL, aND haNs zima the Rise and fall of high Performance fortran

received by the community, generating a great deal of interest in the language and its implementation, but never succeeded in building a strong user base. We concentrate on the technical and social process of designing HPF, as well as the factors that contributed to its successes and its shortcomings.

HPF was one of many efforts to develop parallel programming techniques active since the late 1980s. Here, we do not treat these efforts comprehensively or review all work prior to HPF; more on these topics can be found in the conference version of this article15 and in the History of Programming Languages conferences (http:// www.acm.org/sigplan/hopl). Rather, we provide background on the parallel computers and programming systems available when HPF was developed, then turn to the process of defining HPF, the features of the language itself, and its successes and failures. We conclude with a discussion of lessons to be learned by developers and users of other languages.

 

PArALLeLiSm—the exeCUtion of multiple tasks at the same time—is fundamental in computer design. even very early computer systems employed low-level parallelism (such as overlapping input-output with computing). Later designs moved parallelism to higher levels, including vector processors in the 1970s and shared-memory and distributed-memory parallel supercomputers in the 1980s. Software support for these machines never fully reached the standard of commercial software development, due mainly to the specialized market for high-performance computing and the difficulty related to users’ extreme focus on target-code performance. Here, we trace the history of High Performance Fortran (HPF), a high-level data-parallel language standardized by a group of industry and academic leaders, 1991–1993. It was initially well

Parallel computers and Languages The milieu in which HPF was created (see the figure here) guides much of our discussion, with rectangles representing classes of computers and attached circles representing the most common programming languages for

 

key insights

;;; Programming high-performance parallel computing systems has been dominated by a low-level approach requiring explicit management of communication; hPf elevated parallel programming to a higher level of abstraction, enhancing portability, programmer productivity, and dependability.

;;; many ideas pioneered by hPf (such as high-level data distribution directives and parallel loop constructs) have significantly influenced modern parallel languages, including chapel, X- 10, and fortress.

;;; hPf’s expressive power led to early success, but program performance could not always match that of mPi-based programs, resulting in failure to win over a broad user community.